Related
8 bits representing the number 7 look like this:
00000111
Three bits are set.
What are the algorithms to determine the number of set bits in a 32-bit integer?
This is known as the 'Hamming Weight', 'popcount' or 'sideways addition'.
Some CPUs have a single built-in instruction to do it and others have parallel instructions which act on bit vectors. Instructions like x86's popcnt (on CPUs where it's supported) will almost certainly be fastest for a single integer. Some other architectures may have a slow instruction implemented with a microcoded loop that tests a bit per cycle (citation needed - hardware popcount is normally fast if it exists at all.).
The 'best' algorithm really depends on which CPU you are on and what your usage pattern is.
Your compiler may know how to do something that's good for the specific CPU you're compiling for, e.g. C++20 std::popcount(), or C++ std::bitset<32>::count(), as a portable way to access builtin / intrinsic functions (see another answer on this question). But your compiler's choice of fallback for target CPUs that don't have hardware popcnt might not be optimal for your use-case. Or your language (e.g. C) might not expose any portable function that could use a CPU-specific popcount when there is one.
Portable algorithms that don't need (or benefit from) any HW support
A pre-populated table lookup method can be very fast if your CPU has a large cache and you are doing lots of these operations in a tight loop. However it can suffer because of the expense of a 'cache miss', where the CPU has to fetch some of the table from main memory. (Look up each byte separately to keep the table small.) If you want popcount for a contiguous range of numbers, only the low byte is changing for groups of 256 numbers, making this very good.
If you know that your bytes will be mostly 0's or mostly 1's then there are efficient algorithms for these scenarios, e.g. clearing the lowest set with a bithack in a loop until it becomes zero.
I believe a very good general purpose algorithm is the following, known as 'parallel' or 'variable-precision SWAR algorithm'. I have expressed this in a C-like pseudo language, you may need to adjust it to work for a particular language (e.g. using uint32_t for C++ and >>> in Java):
GCC10 and clang 10.0 can recognize this pattern / idiom and compile it to a hardware popcnt or equivalent instruction when available, giving you the best of both worlds. (https://godbolt.org/z/qGdh1dvKK)
int numberOfSetBits(uint32_t i)
{
// Java: use int, and use >>> instead of >>. Or use Integer.bitCount()
// C or C++: use uint32_t
i = i - ((i >> 1) & 0x55555555); // add pairs of bits
i = (i & 0x33333333) + ((i >> 2) & 0x33333333); // quads
i = (i + (i >> 4)) & 0x0F0F0F0F; // groups of 8
return (i * 0x01010101) >> 24; // horizontal sum of bytes
}
For JavaScript: coerce to integer with |0 for performance: change the first line to i = (i|0) - ((i >> 1) & 0x55555555);
This has the best worst-case behaviour of any of the algorithms discussed, so will efficiently deal with any usage pattern or values you throw at it. (Its performance is not data-dependent on normal CPUs where all integer operations including multiply are constant-time. It doesn't get any faster with "simple" inputs, but it's still pretty decent.)
References:
https://graphics.stanford.edu/~seander/bithacks.html
https://catonmat.net/low-level-bit-hacks for bithack basics, like how subtracting 1 flips contiguous zeros.
https://en.wikipedia.org/wiki/Hamming_weight
http://gurmeet.net/puzzles/fast-bit-counting-routines/
http://aggregate.ee.engr.uky.edu/MAGIC/#Population%20Count%20(Ones%20Count)
How this SWAR bithack works:
i = i - ((i >> 1) & 0x55555555);
The first step is an optimized version of masking to isolate the odd / even bits, shifting to line them up, and adding. This effectively does 16 separate additions in 2-bit accumulators (SWAR = SIMD Within A Register). Like (i & 0x55555555) + ((i>>1) & 0x55555555).
The next step takes the odd/even eight of those 16x 2-bit accumulators and adds again, producing 8x 4-bit sums. The i - ... optimization isn't possible this time so it does just mask before / after shifting. Using the same 0x33... constant both times instead of 0xccc... before shifting is a good thing when compiling for ISAs that need to construct 32-bit constants in registers separately.
The final shift-and-add step of (i + (i >> 4)) & 0x0F0F0F0F widens to 4x 8-bit accumulators. It masks after adding instead of before, because the maximum value in any 4-bit accumulator is 4, if all 4 bits of the corresponding input bits were set. 4+4 = 8 which still fits in 4 bits, so carry between nibble elements is impossible in i + (i >> 4).
So far this is just fairly normal SIMD using SWAR techniques with a few clever optimizations. Continuing on with the same pattern for 2 more steps can widen to 2x 16-bit then 1x 32-bit counts. But there is a more efficient way on machines with fast hardware multiply:
Once we have few enough "elements", a multiply with a magic constant can sum all the elements into the top element. In this case byte elements. Multiply is done by left-shifting and adding, so a multiply of x * 0x01010101 results in x + (x<<8) + (x<<16) + (x<<24). Our 8-bit elements are wide enough (and holding small enough counts) that this doesn't produce carry into that top 8 bits.
A 64-bit version of this can do 8x 8-bit elements in a 64-bit integer with a 0x0101010101010101 multiplier, and extract the high byte with >>56. So it doesn't take any extra steps, just wider constants. This is what GCC uses for __builtin_popcountll on x86 systems when the hardware popcnt instruction isn't enabled. If you can use builtins or intrinsics for this, do so to give the compiler a chance to do target-specific optimizations.
With full SIMD for wider vectors (e.g. counting a whole array)
This bitwise-SWAR algorithm could parallelize to be done in multiple vector elements at once, instead of in a single integer register, for a speedup on CPUs with SIMD but no usable popcount instruction. (e.g. x86-64 code that has to run on any CPU, not just Nehalem or later.)
However, the best way to use vector instructions for popcount is usually by using a variable-shuffle to do a table-lookup for 4 bits at a time of each byte in parallel. (The 4 bits index a 16 entry table held in a vector register).
On Intel CPUs, the hardware 64bit popcnt instruction can outperform an SSSE3 PSHUFB bit-parallel implementation by about a factor of 2, but only if your compiler gets it just right. Otherwise SSE can come out significantly ahead. Newer compiler versions are aware of the popcnt false dependency problem on Intel.
https://github.com/WojciechMula/sse-popcount state-of-the-art x86 SIMD popcount for SSSE3, AVX2, AVX512BW, AVX512VBMI, or AVX512 VPOPCNT. Using Harley-Seal across vectors to defer popcount within an element. (Also ARM NEON)
Counting 1 bits (population count) on large data using AVX-512 or AVX-2
related: https://github.com/mklarqvist/positional-popcount - separate counts for each bit-position of multiple 8, 16, 32, or 64-bit integers. (Again, x86 SIMD including AVX-512 which is really good at this, with vpternlogd making Harley-Seal very good.)
Some languages portably expose the operation in a way that can use efficient hardware support if available, otherwise some library fallback that's hopefully decent.
For example (from a table by language):
C++ has std::bitset<>::count(), or C++20 std::popcount(T x)
Java has java.lang.Integer.bitCount() (also for Long or BigInteger)
C# has System.Numerics.BitOperations.PopCount()
Python has int.bit_count() (since 3.10)
Not all compilers / libraries actually manage to use HW support when it's available, though. (Notably MSVC, even with options that make std::popcount inline as x86 popcnt, its std::bitset::count still always uses a lookup table. This will hopefully change in future versions.)
Also consider the built-in functions of your compiler when the portable language doesn't have this basic bit operation. In GNU C for example:
int __builtin_popcount (unsigned int x);
int __builtin_popcountll (unsigned long long x);
In the worst case (no single-instruction HW support) the compiler will generate a call to a function (which in current GCC uses a shift/and bit-hack like this answer, at least for x86). In the best case the compiler will emit a cpu instruction to do the job. (Just like a * or / operator - GCC will use a hardware multiply or divide instruction if available, otherwise will call a libgcc helper function.) Or even better, if the operand is a compile-time constant after inlining, it can do constant-propagation to get a compile-time-constant popcount result.
The GCC builtins even work across multiple platforms. Popcount has almost become mainstream in the x86 architecture, so it makes sense to start using the builtin now so you can recompile to let it inline a hardware instruction when you compile with -mpopcnt or something that includes that (e.g. https://godbolt.org/z/Ma5e5a). Other architectures have had popcount for years, but in the x86 world there are still some ancient Core 2 and similar vintage AMD CPUs in use.
On x86, you can tell the compiler that it can assume support for popcnt instruction with -mpopcnt (also implied by -msse4.2). See GCC x86 options. -march=nehalem -mtune=skylake (or -march= whatever CPU you want your code to assume and to tune for) could be a good choice. Running the resulting binary on an older CPU will result in an illegal-instruction fault.
To make binaries optimized for the machine you build them on, use -march=native (with gcc, clang, or ICC).
MSVC provides an intrinsic for the x86 popcnt instruction, but unlike gcc it's really an intrinsic for the hardware instruction and requires hardware support.
Using std::bitset<>::count() instead of a built-in
In theory, any compiler that knows how to popcount efficiently for the target CPU should expose that functionality through ISO C++ std::bitset<>. In practice, you might be better off with the bit-hack AND/shift/ADD in some cases for some target CPUs.
For target architectures where hardware popcount is an optional extension (like x86), not all compilers have a std::bitset that takes advantage of it when available. For example, MSVC has no way to enable popcnt support at compile time, and it's std::bitset<>::count always uses a table lookup, even with /Ox /arch:AVX (which implies SSE4.2, which in turn implies the popcnt feature.) (Update: see below; that does get MSVC's C++20 std::popcount to use x86 popcnt, but still not its bitset<>::count. MSVC could fix that by updating their standard library headers to use std::popcount when available.)
But at least you get something portable that works everywhere, and with gcc/clang with the right target options, you get hardware popcount for architectures that support it.
#include <bitset>
#include <limits>
#include <type_traits>
template<typename T>
//static inline // static if you want to compile with -mpopcnt in one compilation unit but not others
typename std::enable_if<std::is_integral<T>::value, unsigned >::type
popcount(T x)
{
static_assert(std::numeric_limits<T>::radix == 2, "non-binary type");
// sizeof(x)*CHAR_BIT
constexpr int bitwidth = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
// std::bitset constructor was only unsigned long before C++11. Beware if porting to C++03
static_assert(bitwidth <= std::numeric_limits<unsigned long long>::digits, "arg too wide for std::bitset() constructor");
typedef typename std::make_unsigned<T>::type UT; // probably not needed, bitset width chops after sign-extension
std::bitset<bitwidth> bs( static_cast<UT>(x) );
return bs.count();
}
See asm from gcc, clang, icc, and MSVC on the Godbolt compiler explorer.
x86-64 gcc -O3 -std=gnu++11 -mpopcnt emits this:
unsigned test_short(short a) { return popcount(a); }
movzx eax, di # note zero-extension, not sign-extension
popcnt rax, rax
ret
unsigned test_int(int a) { return popcount(a); }
mov eax, edi
popcnt rax, rax # unnecessary 64-bit operand size
ret
unsigned test_u64(unsigned long long a) { return popcount(a); }
xor eax, eax # gcc avoids false dependencies for Intel CPUs
popcnt rax, rdi
ret
PowerPC64 gcc -O3 -std=gnu++11 emits (for the int arg version):
rldicl 3,3,0,32 # zero-extend from 32 to 64-bit
popcntd 3,3 # popcount
blr
This source isn't x86-specific or GNU-specific at all, but only compiles well with gcc/clang/icc, at least when targeting x86 (including x86-64).
Also note that gcc's fallback for architectures without single-instruction popcount is a byte-at-a-time table lookup. This isn't wonderful for ARM, for example.
C++20 has std::popcount(T)
Current libstdc++ headers unfortunately define it with a special case if(x==0) return 0; at the start, which clang doesn't optimize away when compiling for x86:
#include <bit>
int bar(unsigned x) {
return std::popcount(x);
}
clang 11.0.1 -O3 -std=gnu++20 -march=nehalem (https://godbolt.org/z/arMe5a)
# clang 11
bar(unsigned int): # #bar(unsigned int)
popcnt eax, edi
cmove eax, edi # redundant: if popcnt result is 0, return the original 0 instead of the popcnt-generated 0...
ret
But GCC compiles nicely:
# gcc 10
xor eax, eax # break false dependency on Intel SnB-family before Ice Lake.
popcnt eax, edi
ret
Even MSVC does well with it, as long as you use -arch:AVX or later (and enable C++20 with -std:c++latest). https://godbolt.org/z/7K4Gef
int bar(unsigned int) PROC ; bar, COMDAT
popcnt eax, ecx
ret 0
int bar(unsigned int) ENDP ; bar
In my opinion, the "best" solution is the one that can be read by another programmer (or the original programmer two years later) without copious comments. You may well want the fastest or cleverest solution which some have already provided but I prefer readability over cleverness any time.
unsigned int bitCount (unsigned int value) {
unsigned int count = 0;
while (value > 0) { // until all bits are zero
if ((value & 1) == 1) // check lower bit
count++;
value >>= 1; // shift bits, removing lower bit
}
return count;
}
If you want more speed (and assuming you document it well to help out your successors), you could use a table lookup:
// Lookup table for fast calculation of bits set in 8-bit unsigned char.
static unsigned char oneBitsInUChar[] = {
// 0 1 2 3 4 5 6 7 8 9 A B C D E F (<- n)
// =====================================================
0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, // 0n
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, // 1n
: : :
4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8, // Fn
};
// Function for fast calculation of bits set in 16-bit unsigned short.
unsigned char oneBitsInUShort (unsigned short x) {
return oneBitsInUChar [x >> 8]
+ oneBitsInUChar [x & 0xff];
}
// Function for fast calculation of bits set in 32-bit unsigned int.
unsigned char oneBitsInUInt (unsigned int x) {
return oneBitsInUShort (x >> 16)
+ oneBitsInUShort (x & 0xffff);
}
These rely on specific data type sizes so they're not that portable. But, since many performance optimisations aren't portable anyway, that may not be an issue. If you want portability, I'd stick to the readable solution.
From Hacker's Delight, p. 66, Figure 5-2
int pop(unsigned x)
{
x = x - ((x >> 1) & 0x55555555);
x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
x = (x + (x >> 4)) & 0x0F0F0F0F;
x = x + (x >> 8);
x = x + (x >> 16);
return x & 0x0000003F;
}
Executes in ~20-ish instructions (arch dependent), no branching.Hacker's Delight is delightful! Highly recommended.
I think the fastest way—without using lookup tables and popcount—is the following. It counts the set bits with just 12 operations.
int popcount(int v) {
v = v - ((v >> 1) & 0x55555555); // put count of each 2 bits into those 2 bits
v = (v & 0x33333333) + ((v >> 2) & 0x33333333); // put count of each 4 bits into those 4 bits
return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
}
It works because you can count the total number of set bits by dividing in two halves, counting the number of set bits in both halves and then adding them up. Also know as Divide and Conquer paradigm. Let's get into detail..
v = v - ((v >> 1) & 0x55555555);
The number of bits in two bits can be 0b00, 0b01 or 0b10. Lets try to work this out on 2 bits..
---------------------------------------------
| v | (v >> 1) & 0b0101 | v - x |
---------------------------------------------
0b00 0b00 0b00
0b01 0b00 0b01
0b10 0b01 0b01
0b11 0b01 0b10
This is what was required: the last column shows the count of set bits in every two bit pair. If the two bit number is >= 2 (0b10) then and produces 0b01, else it produces 0b00.
v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
This statement should be easy to understand. After the first operation we have the count of set bits in every two bits, now we sum up that count in every 4 bits.
v & 0b00110011 //masks out even two bits
(v >> 2) & 0b00110011 // masks out odd two bits
We then sum up the above result, giving us the total count of set bits in 4 bits. The last statement is the most tricky.
c = ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
Let's break it down further...
v + (v >> 4)
It's similar to the second statement; we are counting the set bits in groups of 4 instead. We know—because of our previous operations—that every nibble has the count of set bits in it. Let's look an example. Suppose we have the byte 0b01000010. It means the first nibble has its 4bits set and the second one has its 2bits set. Now we add those nibbles together.
v = 0b01000010
(v >> 4) = 0b00000100
v + (v >> 4) = 0b01000010 + 0b00000100
It gives us the count of set bits in a byte, in the second nibble 0b01000110 and therefore we mask the first four bytes of all the bytes in the number (discarding them).
0b01000110 & 0x0F = 0b00000110
Now every byte has the count of set bits in it. We need to add them up all together. The trick is to multiply the result by 0b10101010 which has an interesting property. If our number has four bytes, A B C D, it will result in a new number with these bytes A+B+C+D B+C+D C+D D. A 4 byte number can have maximum of 32 bits set, which can be represented as 0b00100000.
All we need now is the first byte which has the sum of all set bits in all the bytes, and we get it by >> 24. This algorithm was designed for 32 bit words but can be easily modified for 64 bit words.
If you happen to be using Java, the built-in method Integer.bitCount will do that.
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.
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)
+-------------------------------+
This is one of those questions where it helps to know your micro-architecture. I just timed two variants under gcc 4.3.3 compiled with -O3 using C++ inlines to eliminate function call overhead, one billion iterations, keeping the running sum of all counts to ensure the compiler doesn't remove anything important, using rdtsc for timing (clock cycle precise).
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;
}
The unmodified Hacker's Delight took 12.2 gigacycles. My parallel version (counting twice as many bits) runs in 13.0 gigacycles. 10.5s total elapsed for both together on a 2.4GHz Core Duo. 25 gigacycles = just over 10 seconds at this clock frequency, so I'm confident my timings are right.
This has to do with instruction dependency chains, which are very bad for this algorithm. I could nearly double the speed again by using a pair of 64-bit registers. In fact, if I was clever and added x+y a little sooner I could shave off some shifts. The 64-bit version with some small tweaks would come out about even, but count twice as many bits again.
With 128 bit SIMD registers, yet another factor of two, and the SSE instruction sets often have clever short-cuts, too.
There's no reason for the code to be especially transparent. The interface is simple, the algorithm can be referenced on-line in many places, and it's amenable to comprehensive unit test. The programmer who stumbles upon it might even learn something. These bit operations are extremely natural at the machine level.
OK, I decided to bench the tweaked 64-bit version. For this one 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;
}
That looks about right (I'm not testing carefully, though). Now the timings come out at 10.70 gigacycles / 14.1 gigacycles. That later number summed 128 billion bits and corresponds to 5.9s elapsed on this machine. The non-parallel version speeds up a tiny bit because I'm running in 64-bit mode and it likes 64-bit registers slightly better than 32-bit registers.
Let's see if there's a bit more OOO pipelining to be had here. This was a bit more involved, so I actually tested a bit. Each term alone sums to 64, all combined sum to 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;
}
I was excited for a moment, but it turns out gcc is playing inline tricks with -O3 even though I'm not using the inline keyword in some tests. When I let gcc play tricks, a billion calls to pop4() takes 12.56 gigacycles, but I determined it was folding arguments as constant expressions. A more realistic number appears to be 19.6gc for another 30% speed-up. My test loop now looks like this, making sure each argument is different enough to stop gcc from playing tricks.
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();
256 billion bits summed in 8.17s elapsed. Works out to 1.02s for 32 million bits as benchmarked in the 16-bit table lookup. Can't compare directly, because the other bench doesn't give a clock speed, but looks like I've slapped the snot out of the 64KB table edition, which is a tragic use of L1 cache in the first place.
Update: decided to do the obvious and create pop6() by adding four more duplicated lines. Came out to 22.8gc, 384 billion bits summed in 9.5s elapsed. So there's another 20% Now at 800ms for 32 billion bits.
Why not iteratively divide by 2?
count = 0
while n > 0
if (n % 2) == 1
count += 1
n /= 2
I agree that this isn't the fastest, but "best" is somewhat ambiguous. I'd argue though that "best" should have an element of clarity
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.
For a happy medium between a 232 lookup table and iterating through each bit individually:
int bitcount(unsigned int num){
int count = 0;
static int nibblebits[] =
{0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4};
for(; num != 0; num >>= 4)
count += nibblebits[num & 0x0f];
return count;
}
From http://ctips.pbwiki.com/CountBits
This can be done in O(k), where k is the number of bits set.
int NumberOfSetBits(int n)
{
int count = 0;
while (n){
++ count;
n = (n - 1) & n;
}
return count;
}
It's not the fastest or best solution, but I found the same question in my way, and I started to think and think. finally I realized that it can be done like this if you get the problem from mathematical side, and draw a graph, then you find that it's a function which has some periodic part, and then you realize the difference between the periods... so here you go:
unsigned int f(unsigned int x)
{
switch (x) {
case 0:
return 0;
case 1:
return 1;
case 2:
return 1;
case 3:
return 2;
default:
return f(x/4) + f(x%4);
}
}
I think the Brian Kernighan's method will be useful too...
It goes through as many iterations as there are set bits. So if we have a 32-bit word with only the high bit set, then it will only go once through the loop.
int countSetBits(unsigned int n) {
unsigned int n; // count the number of bits set in n
unsigned int c; // c accumulates the total bits set in n
for (c=0;n>0;n=n&(n-1)) c++;
return c;
}
Published in 1988, the C Programming Language 2nd Ed. (by Brian W. Kernighan and Dennis M. Ritchie) mentions this in exercise 2-9. On April 19, 2006 Don Knuth pointed out to me that this method "was first published by Peter Wegner in CACM 3 (1960), 322. (Also discovered independently by Derrick Lehmer and published in 1964 in a book edited by Beckenbach.)"
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.
private int get_bits_set(int v)
{
int c; // 'c' accumulates the total bits set in 'v'
for (c = 0; v>0; c++)
{
v &= v - 1; // Clear the least significant bit set
}
return c;
}
Few open questions:-
If the number is negative then?
If the number is 1024 , then the "iteratively divide by 2" method will iterate 10 times.
we can modify the algo to support the negative number as follows:-
count = 0
while n != 0
if ((n % 2) == 1 || (n % 2) == -1
count += 1
n /= 2
return count
now to overcome the second problem we can write the algo like:-
int bit_count(int num)
{
int count=0;
while(num)
{
num=(num)&(num-1);
count++;
}
return count;
}
for complete reference see :
http://goursaha.freeoda.com/Miscellaneous/IntegerBitCount.html
I use the below code which is more intuitive.
int countSetBits(int n) {
return !n ? 0 : 1 + countSetBits(n & (n-1));
}
Logic : n & (n-1) resets the last set bit of n.
P.S : I know this is not O(1) solution, albeit an interesting solution.
What do you means with "Best algorithm"? The shorted code or the fasted code? Your code look very elegant and it has a constant execution time. The code is also very short.
But if the speed is the major factor and not the code size then I think the follow can be faster:
static final int[] BIT_COUNT = { 0, 1, 1, ... 256 values with a bitsize of a byte ... };
static int bitCountOfByte( int value ){
return BIT_COUNT[ value & 0xFF ];
}
static int bitCountOfInt( int value ){
return bitCountOfByte( value )
+ bitCountOfByte( value >> 8 )
+ bitCountOfByte( value >> 16 )
+ bitCountOfByte( value >> 24 );
}
I think that this will not more faster for a 64 bit value but a 32 bit value can be faster.
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
if you're using C++ another option is to use template metaprogramming:
// recursive template to sum bits in an int
template <int BITS>
int countBits(int val) {
// return the least significant bit plus the result of calling ourselves with
// .. the shifted value
return (val & 0x1) + countBits<BITS-1>(val >> 1);
}
// template specialisation to terminate the recursion when there's only one bit left
template<>
int countBits<1>(int val) {
return val & 0x1;
}
usage would be:
// to count bits in a byte/char (this returns 8)
countBits<8>( 255 )
// another byte (this returns 7)
countBits<8>( 254 )
// counting bits in a word/short (this returns 1)
countBits<16>( 256 )
you could of course further expand this template to use different types (even auto-detecting bit size) but I've kept it simple for clarity.
edit: forgot to mention this is good because it should work in any C++ compiler and it basically just unrolls your loop for you if a constant value is used for the bit count (in other words, I'm pretty sure it's the fastest general method you'll find)
C++20 std::popcount
The following proposal has been merged http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0553r4.html and should add it to a the <bit> header.
I expect the usage to be like:
#include <bit>
#include <iostream>
int main() {
std::cout << std::popcount(0x55) << std::endl;
}
I'll give it a try when support arrives to GCC, GCC 9.1.0 with g++-9 -std=c++2a still doesn't support it.
The proposal says:
Header: <bit>
namespace std {
// 25.5.6, counting
template<class T>
constexpr int popcount(T x) noexcept;
and:
template<class T>
constexpr int popcount(T x) noexcept;
Constraints: T is an unsigned integer type (3.9.1 [basic.fundamental]).
Returns: The number of 1 bits in the value of x.
std::rotl and std::rotr were also added to do circular bit rotations: Best practices for circular shift (rotate) operations in C++
You can do:
while(n){
n = n & (n-1);
count++;
}
The logic behind this is the bits of n-1 is inverted from rightmost set bit of n.
If n=6, i.e., 110 then 5 is 101 the bits are inverted from rightmost set bit of n.
So if we & these two we will make the rightmost bit 0 in every iteration and always go to the next rightmost set bit. Hence, counting the set bit. The worst time complexity will be O(log n) when every bit is set.
I'm particularly fond of this example from the fortune file:
#define BITCOUNT(x) (((BX_(x)+(BX_(x)>>4)) & 0x0F0F0F0F) % 255)
#define BX_(x) ((x) - (((x)>>1)&0x77777777)
- (((x)>>2)&0x33333333)
- (((x)>>3)&0x11111111))
I like it best because it's so pretty!
Java JDK1.5
Integer.bitCount(n);
where n is the number whose 1's are to be counted.
check also,
Integer.highestOneBit(n);
Integer.lowestOneBit(n);
Integer.numberOfLeadingZeros(n);
Integer.numberOfTrailingZeros(n);
//Beginning with the value 1, rotate left 16 times
n = 1;
for (int i = 0; i < 16; i++) {
n = Integer.rotateLeft(n, 1);
System.out.println(n);
}
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];
}
A fast C# solution using a pre-calculated table of Byte bit counts with branching on the input size.
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
};
}
I always use this in competitive programming, and it's easy to write and is efficient:
#include <bits/stdc++.h>
using namespace std;
int countOnes(int n) {
bitset<32> b(n);
return b.count();
}
There are many algorithm to count the set bits; but i think the best one is the faster one!
You can see the detailed on this page:
Bit Twiddling Hacks
I suggest this one:
Counting bits set in 14, 24, or 32-bit words using 64-bit instructions
unsigned int v; // count the number of bits set in v
unsigned int c; // c accumulates the total bits set in v
// option 1, for at most 14-bit values in v:
c = (v * 0x200040008001ULL & 0x111111111111111ULL) % 0xf;
// option 2, for at most 24-bit values in v:
c = ((v & 0xfff) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
c += (((v & 0xfff000) >> 12) * 0x1001001001001ULL & 0x84210842108421ULL)
% 0x1f;
// option 3, for at most 32-bit values in v:
c = ((v & 0xfff) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
c += (((v & 0xfff000) >> 12) * 0x1001001001001ULL & 0x84210842108421ULL) %
0x1f;
c += ((v >> 24) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
This method requires a 64-bit CPU with fast modulus division to be efficient. The first option takes only 3 operations; the second option takes 10; and the third option takes 15.
I need to merge two variables into one like in example below:
x = 0b110101
y = 0b10011
merged(x,y) -> 0b11010111001
merged(y,x) -> 0b10011101011
So, y is reversed and concatenated to x without unnecessary zeros (last bit of result is always 1)
In other words: merged(abcdefg, hijklm) results in abcdefgmlkjih where a and h are 1
What would be the most efficient way to do this in C
EDIT: most efficient = fastest, sizeof(x) = sizeof(y) = 1, CHAR_BIT= 8
EDIT2: Since question has been put on hold I will post summary right here:
Further limitations and more detail:
Target language: C with 64-bit operations support
For this exact problem fastest method would be a 256 element lookup table as was suggested in comments, which returns index of high bit N and reversed value of second argument, so we can shift first argument to the left by N and perform bitwise or with reversed value of second argument.
In case someone needs good performance for arguments larger than 8 bit (lookup table is not an option) they could:
Find index of high bit using de Bruijn algorithm. Example for 32 bits:
uint32_t msbDeBruijn32( uint32_t v )
{
static const int MultiplyDeBruijnBitPosition[32] =
{
0, 9, 1, 10, 13, 21, 2, 29, 11, 14, 16, 18, 22, 25, 3, 30,
8, 12, 20, 28, 15, 17, 24, 7, 19, 27, 23, 6, 26, 5, 4, 31
};
v |= v >> 1; // first round down to one less than a power of 2
v |= v >> 2;
v |= v >> 4;
v |= v >> 8;
v |= v >> 16;
return MultiplyDeBruijnBitPosition[( uint32_t )( v * 0x07C4ACDDU ) >> 27];
}
taken from this question:
Find most significant bit (left-most) that is set in a bit array
Reverse bits in bytes (one by one) of the second argument using this bit twiddling hack
unsigned char b; // reverse this (8-bit) byte
b = (b * 0x0202020202ULL & 0x010884422010ULL) % 1023;
taken from here: http://graphics.stanford.edu/~seander/bithacks.html#BitReverseObvious
So, this question now can be closed/deleted
Solution
This should do the trick in C++ and C:
unsigned int merged(unsigned char a, unsigned char b) {
unsigned int r = a;
for (; b; b>>=1)
r = (r<<1)|(b&1);
return r;
}
I don't think that there's a more efficient way to do it. Here an online demo.
How does it work ?
This works by building the result by starting with the first argument:
r is 0b110101;
Then it left-shifts r, which is, using your terminology, like concatenating a 0 at the lowest end of r:
0b110101 << 1 is 0b1101010;
and simultaneously we get the lowest bit of b with a bitwise and, then set the lowest bit of r to the same value using a bitwise or:
0b10011 & 1 is 0b00001
so (0b110101 << 1)|(0b10011 & 1) is 0b1101011;
We then right-shift b to process the next lowest bit in the same way:
0b10011 >> 1 is 0b1001;
as long as there is some bit set left in b.
Important remarks
You have to be careful about the type used in order to avoid the risk of overflow. So for 2 unsigned char as input you could use unsigned int as output.
Note that using signed integers would be risky, if the left-shift might cause the sign bit to change. This is why unsigned is important here.
8 bits representing the number 7 look like this:
00000111
Three bits are set.
What are the algorithms to determine the number of set bits in a 32-bit integer?
This is known as the 'Hamming Weight', 'popcount' or 'sideways addition'.
Some CPUs have a single built-in instruction to do it and others have parallel instructions which act on bit vectors. Instructions like x86's popcnt (on CPUs where it's supported) will almost certainly be fastest for a single integer. Some other architectures may have a slow instruction implemented with a microcoded loop that tests a bit per cycle (citation needed - hardware popcount is normally fast if it exists at all.).
The 'best' algorithm really depends on which CPU you are on and what your usage pattern is.
Your compiler may know how to do something that's good for the specific CPU you're compiling for, e.g. C++20 std::popcount(), or C++ std::bitset<32>::count(), as a portable way to access builtin / intrinsic functions (see another answer on this question). But your compiler's choice of fallback for target CPUs that don't have hardware popcnt might not be optimal for your use-case. Or your language (e.g. C) might not expose any portable function that could use a CPU-specific popcount when there is one.
Portable algorithms that don't need (or benefit from) any HW support
A pre-populated table lookup method can be very fast if your CPU has a large cache and you are doing lots of these operations in a tight loop. However it can suffer because of the expense of a 'cache miss', where the CPU has to fetch some of the table from main memory. (Look up each byte separately to keep the table small.) If you want popcount for a contiguous range of numbers, only the low byte is changing for groups of 256 numbers, making this very good.
If you know that your bytes will be mostly 0's or mostly 1's then there are efficient algorithms for these scenarios, e.g. clearing the lowest set with a bithack in a loop until it becomes zero.
I believe a very good general purpose algorithm is the following, known as 'parallel' or 'variable-precision SWAR algorithm'. I have expressed this in a C-like pseudo language, you may need to adjust it to work for a particular language (e.g. using uint32_t for C++ and >>> in Java):
GCC10 and clang 10.0 can recognize this pattern / idiom and compile it to a hardware popcnt or equivalent instruction when available, giving you the best of both worlds. (https://godbolt.org/z/qGdh1dvKK)
int numberOfSetBits(uint32_t i)
{
// Java: use int, and use >>> instead of >>. Or use Integer.bitCount()
// C or C++: use uint32_t
i = i - ((i >> 1) & 0x55555555); // add pairs of bits
i = (i & 0x33333333) + ((i >> 2) & 0x33333333); // quads
i = (i + (i >> 4)) & 0x0F0F0F0F; // groups of 8
return (i * 0x01010101) >> 24; // horizontal sum of bytes
}
For JavaScript: coerce to integer with |0 for performance: change the first line to i = (i|0) - ((i >> 1) & 0x55555555);
This has the best worst-case behaviour of any of the algorithms discussed, so will efficiently deal with any usage pattern or values you throw at it. (Its performance is not data-dependent on normal CPUs where all integer operations including multiply are constant-time. It doesn't get any faster with "simple" inputs, but it's still pretty decent.)
References:
https://graphics.stanford.edu/~seander/bithacks.html
https://catonmat.net/low-level-bit-hacks for bithack basics, like how subtracting 1 flips contiguous zeros.
https://en.wikipedia.org/wiki/Hamming_weight
http://gurmeet.net/puzzles/fast-bit-counting-routines/
http://aggregate.ee.engr.uky.edu/MAGIC/#Population%20Count%20(Ones%20Count)
How this SWAR bithack works:
i = i - ((i >> 1) & 0x55555555);
The first step is an optimized version of masking to isolate the odd / even bits, shifting to line them up, and adding. This effectively does 16 separate additions in 2-bit accumulators (SWAR = SIMD Within A Register). Like (i & 0x55555555) + ((i>>1) & 0x55555555).
The next step takes the odd/even eight of those 16x 2-bit accumulators and adds again, producing 8x 4-bit sums. The i - ... optimization isn't possible this time so it does just mask before / after shifting. Using the same 0x33... constant both times instead of 0xccc... before shifting is a good thing when compiling for ISAs that need to construct 32-bit constants in registers separately.
The final shift-and-add step of (i + (i >> 4)) & 0x0F0F0F0F widens to 4x 8-bit accumulators. It masks after adding instead of before, because the maximum value in any 4-bit accumulator is 4, if all 4 bits of the corresponding input bits were set. 4+4 = 8 which still fits in 4 bits, so carry between nibble elements is impossible in i + (i >> 4).
So far this is just fairly normal SIMD using SWAR techniques with a few clever optimizations. Continuing on with the same pattern for 2 more steps can widen to 2x 16-bit then 1x 32-bit counts. But there is a more efficient way on machines with fast hardware multiply:
Once we have few enough "elements", a multiply with a magic constant can sum all the elements into the top element. In this case byte elements. Multiply is done by left-shifting and adding, so a multiply of x * 0x01010101 results in x + (x<<8) + (x<<16) + (x<<24). Our 8-bit elements are wide enough (and holding small enough counts) that this doesn't produce carry into that top 8 bits.
A 64-bit version of this can do 8x 8-bit elements in a 64-bit integer with a 0x0101010101010101 multiplier, and extract the high byte with >>56. So it doesn't take any extra steps, just wider constants. This is what GCC uses for __builtin_popcountll on x86 systems when the hardware popcnt instruction isn't enabled. If you can use builtins or intrinsics for this, do so to give the compiler a chance to do target-specific optimizations.
With full SIMD for wider vectors (e.g. counting a whole array)
This bitwise-SWAR algorithm could parallelize to be done in multiple vector elements at once, instead of in a single integer register, for a speedup on CPUs with SIMD but no usable popcount instruction. (e.g. x86-64 code that has to run on any CPU, not just Nehalem or later.)
However, the best way to use vector instructions for popcount is usually by using a variable-shuffle to do a table-lookup for 4 bits at a time of each byte in parallel. (The 4 bits index a 16 entry table held in a vector register).
On Intel CPUs, the hardware 64bit popcnt instruction can outperform an SSSE3 PSHUFB bit-parallel implementation by about a factor of 2, but only if your compiler gets it just right. Otherwise SSE can come out significantly ahead. Newer compiler versions are aware of the popcnt false dependency problem on Intel.
https://github.com/WojciechMula/sse-popcount state-of-the-art x86 SIMD popcount for SSSE3, AVX2, AVX512BW, AVX512VBMI, or AVX512 VPOPCNT. Using Harley-Seal across vectors to defer popcount within an element. (Also ARM NEON)
Counting 1 bits (population count) on large data using AVX-512 or AVX-2
related: https://github.com/mklarqvist/positional-popcount - separate counts for each bit-position of multiple 8, 16, 32, or 64-bit integers. (Again, x86 SIMD including AVX-512 which is really good at this, with vpternlogd making Harley-Seal very good.)
Some languages portably expose the operation in a way that can use efficient hardware support if available, otherwise some library fallback that's hopefully decent.
For example (from a table by language):
C++ has std::bitset<>::count(), or C++20 std::popcount(T x)
Java has java.lang.Integer.bitCount() (also for Long or BigInteger)
C# has System.Numerics.BitOperations.PopCount()
Python has int.bit_count() (since 3.10)
Not all compilers / libraries actually manage to use HW support when it's available, though. (Notably MSVC, even with options that make std::popcount inline as x86 popcnt, its std::bitset::count still always uses a lookup table. This will hopefully change in future versions.)
Also consider the built-in functions of your compiler when the portable language doesn't have this basic bit operation. In GNU C for example:
int __builtin_popcount (unsigned int x);
int __builtin_popcountll (unsigned long long x);
In the worst case (no single-instruction HW support) the compiler will generate a call to a function (which in current GCC uses a shift/and bit-hack like this answer, at least for x86). In the best case the compiler will emit a cpu instruction to do the job. (Just like a * or / operator - GCC will use a hardware multiply or divide instruction if available, otherwise will call a libgcc helper function.) Or even better, if the operand is a compile-time constant after inlining, it can do constant-propagation to get a compile-time-constant popcount result.
The GCC builtins even work across multiple platforms. Popcount has almost become mainstream in the x86 architecture, so it makes sense to start using the builtin now so you can recompile to let it inline a hardware instruction when you compile with -mpopcnt or something that includes that (e.g. https://godbolt.org/z/Ma5e5a). Other architectures have had popcount for years, but in the x86 world there are still some ancient Core 2 and similar vintage AMD CPUs in use.
On x86, you can tell the compiler that it can assume support for popcnt instruction with -mpopcnt (also implied by -msse4.2). See GCC x86 options. -march=nehalem -mtune=skylake (or -march= whatever CPU you want your code to assume and to tune for) could be a good choice. Running the resulting binary on an older CPU will result in an illegal-instruction fault.
To make binaries optimized for the machine you build them on, use -march=native (with gcc, clang, or ICC).
MSVC provides an intrinsic for the x86 popcnt instruction, but unlike gcc it's really an intrinsic for the hardware instruction and requires hardware support.
Using std::bitset<>::count() instead of a built-in
In theory, any compiler that knows how to popcount efficiently for the target CPU should expose that functionality through ISO C++ std::bitset<>. In practice, you might be better off with the bit-hack AND/shift/ADD in some cases for some target CPUs.
For target architectures where hardware popcount is an optional extension (like x86), not all compilers have a std::bitset that takes advantage of it when available. For example, MSVC has no way to enable popcnt support at compile time, and it's std::bitset<>::count always uses a table lookup, even with /Ox /arch:AVX (which implies SSE4.2, which in turn implies the popcnt feature.) (Update: see below; that does get MSVC's C++20 std::popcount to use x86 popcnt, but still not its bitset<>::count. MSVC could fix that by updating their standard library headers to use std::popcount when available.)
But at least you get something portable that works everywhere, and with gcc/clang with the right target options, you get hardware popcount for architectures that support it.
#include <bitset>
#include <limits>
#include <type_traits>
template<typename T>
//static inline // static if you want to compile with -mpopcnt in one compilation unit but not others
typename std::enable_if<std::is_integral<T>::value, unsigned >::type
popcount(T x)
{
static_assert(std::numeric_limits<T>::radix == 2, "non-binary type");
// sizeof(x)*CHAR_BIT
constexpr int bitwidth = std::numeric_limits<T>::digits + std::numeric_limits<T>::is_signed;
// std::bitset constructor was only unsigned long before C++11. Beware if porting to C++03
static_assert(bitwidth <= std::numeric_limits<unsigned long long>::digits, "arg too wide for std::bitset() constructor");
typedef typename std::make_unsigned<T>::type UT; // probably not needed, bitset width chops after sign-extension
std::bitset<bitwidth> bs( static_cast<UT>(x) );
return bs.count();
}
See asm from gcc, clang, icc, and MSVC on the Godbolt compiler explorer.
x86-64 gcc -O3 -std=gnu++11 -mpopcnt emits this:
unsigned test_short(short a) { return popcount(a); }
movzx eax, di # note zero-extension, not sign-extension
popcnt rax, rax
ret
unsigned test_int(int a) { return popcount(a); }
mov eax, edi
popcnt rax, rax # unnecessary 64-bit operand size
ret
unsigned test_u64(unsigned long long a) { return popcount(a); }
xor eax, eax # gcc avoids false dependencies for Intel CPUs
popcnt rax, rdi
ret
PowerPC64 gcc -O3 -std=gnu++11 emits (for the int arg version):
rldicl 3,3,0,32 # zero-extend from 32 to 64-bit
popcntd 3,3 # popcount
blr
This source isn't x86-specific or GNU-specific at all, but only compiles well with gcc/clang/icc, at least when targeting x86 (including x86-64).
Also note that gcc's fallback for architectures without single-instruction popcount is a byte-at-a-time table lookup. This isn't wonderful for ARM, for example.
C++20 has std::popcount(T)
Current libstdc++ headers unfortunately define it with a special case if(x==0) return 0; at the start, which clang doesn't optimize away when compiling for x86:
#include <bit>
int bar(unsigned x) {
return std::popcount(x);
}
clang 11.0.1 -O3 -std=gnu++20 -march=nehalem (https://godbolt.org/z/arMe5a)
# clang 11
bar(unsigned int): # #bar(unsigned int)
popcnt eax, edi
cmove eax, edi # redundant: if popcnt result is 0, return the original 0 instead of the popcnt-generated 0...
ret
But GCC compiles nicely:
# gcc 10
xor eax, eax # break false dependency on Intel SnB-family before Ice Lake.
popcnt eax, edi
ret
Even MSVC does well with it, as long as you use -arch:AVX or later (and enable C++20 with -std:c++latest). https://godbolt.org/z/7K4Gef
int bar(unsigned int) PROC ; bar, COMDAT
popcnt eax, ecx
ret 0
int bar(unsigned int) ENDP ; bar
In my opinion, the "best" solution is the one that can be read by another programmer (or the original programmer two years later) without copious comments. You may well want the fastest or cleverest solution which some have already provided but I prefer readability over cleverness any time.
unsigned int bitCount (unsigned int value) {
unsigned int count = 0;
while (value > 0) { // until all bits are zero
if ((value & 1) == 1) // check lower bit
count++;
value >>= 1; // shift bits, removing lower bit
}
return count;
}
If you want more speed (and assuming you document it well to help out your successors), you could use a table lookup:
// Lookup table for fast calculation of bits set in 8-bit unsigned char.
static unsigned char oneBitsInUChar[] = {
// 0 1 2 3 4 5 6 7 8 9 A B C D E F (<- n)
// =====================================================
0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4, // 0n
1, 2, 2, 3, 2, 3, 3, 4, 2, 3, 3, 4, 3, 4, 4, 5, // 1n
: : :
4, 5, 5, 6, 5, 6, 6, 7, 5, 6, 6, 7, 6, 7, 7, 8, // Fn
};
// Function for fast calculation of bits set in 16-bit unsigned short.
unsigned char oneBitsInUShort (unsigned short x) {
return oneBitsInUChar [x >> 8]
+ oneBitsInUChar [x & 0xff];
}
// Function for fast calculation of bits set in 32-bit unsigned int.
unsigned char oneBitsInUInt (unsigned int x) {
return oneBitsInUShort (x >> 16)
+ oneBitsInUShort (x & 0xffff);
}
These rely on specific data type sizes so they're not that portable. But, since many performance optimisations aren't portable anyway, that may not be an issue. If you want portability, I'd stick to the readable solution.
From Hacker's Delight, p. 66, Figure 5-2
int pop(unsigned x)
{
x = x - ((x >> 1) & 0x55555555);
x = (x & 0x33333333) + ((x >> 2) & 0x33333333);
x = (x + (x >> 4)) & 0x0F0F0F0F;
x = x + (x >> 8);
x = x + (x >> 16);
return x & 0x0000003F;
}
Executes in ~20-ish instructions (arch dependent), no branching.Hacker's Delight is delightful! Highly recommended.
I think the fastest way—without using lookup tables and popcount—is the following. It counts the set bits with just 12 operations.
int popcount(int v) {
v = v - ((v >> 1) & 0x55555555); // put count of each 2 bits into those 2 bits
v = (v & 0x33333333) + ((v >> 2) & 0x33333333); // put count of each 4 bits into those 4 bits
return ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
}
It works because you can count the total number of set bits by dividing in two halves, counting the number of set bits in both halves and then adding them up. Also know as Divide and Conquer paradigm. Let's get into detail..
v = v - ((v >> 1) & 0x55555555);
The number of bits in two bits can be 0b00, 0b01 or 0b10. Lets try to work this out on 2 bits..
---------------------------------------------
| v | (v >> 1) & 0b0101 | v - x |
---------------------------------------------
0b00 0b00 0b00
0b01 0b00 0b01
0b10 0b01 0b01
0b11 0b01 0b10
This is what was required: the last column shows the count of set bits in every two bit pair. If the two bit number is >= 2 (0b10) then and produces 0b01, else it produces 0b00.
v = (v & 0x33333333) + ((v >> 2) & 0x33333333);
This statement should be easy to understand. After the first operation we have the count of set bits in every two bits, now we sum up that count in every 4 bits.
v & 0b00110011 //masks out even two bits
(v >> 2) & 0b00110011 // masks out odd two bits
We then sum up the above result, giving us the total count of set bits in 4 bits. The last statement is the most tricky.
c = ((v + (v >> 4) & 0xF0F0F0F) * 0x1010101) >> 24;
Let's break it down further...
v + (v >> 4)
It's similar to the second statement; we are counting the set bits in groups of 4 instead. We know—because of our previous operations—that every nibble has the count of set bits in it. Let's look an example. Suppose we have the byte 0b01000010. It means the first nibble has its 4bits set and the second one has its 2bits set. Now we add those nibbles together.
v = 0b01000010
(v >> 4) = 0b00000100
v + (v >> 4) = 0b01000010 + 0b00000100
It gives us the count of set bits in a byte, in the second nibble 0b01000110 and therefore we mask the first four bytes of all the bytes in the number (discarding them).
0b01000110 & 0x0F = 0b00000110
Now every byte has the count of set bits in it. We need to add them up all together. The trick is to multiply the result by 0b10101010 which has an interesting property. If our number has four bytes, A B C D, it will result in a new number with these bytes A+B+C+D B+C+D C+D D. A 4 byte number can have maximum of 32 bits set, which can be represented as 0b00100000.
All we need now is the first byte which has the sum of all set bits in all the bytes, and we get it by >> 24. This algorithm was designed for 32 bit words but can be easily modified for 64 bit words.
If you happen to be using Java, the built-in method Integer.bitCount will do that.
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.
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)
+-------------------------------+
This is one of those questions where it helps to know your micro-architecture. I just timed two variants under gcc 4.3.3 compiled with -O3 using C++ inlines to eliminate function call overhead, one billion iterations, keeping the running sum of all counts to ensure the compiler doesn't remove anything important, using rdtsc for timing (clock cycle precise).
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;
}
The unmodified Hacker's Delight took 12.2 gigacycles. My parallel version (counting twice as many bits) runs in 13.0 gigacycles. 10.5s total elapsed for both together on a 2.4GHz Core Duo. 25 gigacycles = just over 10 seconds at this clock frequency, so I'm confident my timings are right.
This has to do with instruction dependency chains, which are very bad for this algorithm. I could nearly double the speed again by using a pair of 64-bit registers. In fact, if I was clever and added x+y a little sooner I could shave off some shifts. The 64-bit version with some small tweaks would come out about even, but count twice as many bits again.
With 128 bit SIMD registers, yet another factor of two, and the SSE instruction sets often have clever short-cuts, too.
There's no reason for the code to be especially transparent. The interface is simple, the algorithm can be referenced on-line in many places, and it's amenable to comprehensive unit test. The programmer who stumbles upon it might even learn something. These bit operations are extremely natural at the machine level.
OK, I decided to bench the tweaked 64-bit version. For this one 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;
}
That looks about right (I'm not testing carefully, though). Now the timings come out at 10.70 gigacycles / 14.1 gigacycles. That later number summed 128 billion bits and corresponds to 5.9s elapsed on this machine. The non-parallel version speeds up a tiny bit because I'm running in 64-bit mode and it likes 64-bit registers slightly better than 32-bit registers.
Let's see if there's a bit more OOO pipelining to be had here. This was a bit more involved, so I actually tested a bit. Each term alone sums to 64, all combined sum to 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;
}
I was excited for a moment, but it turns out gcc is playing inline tricks with -O3 even though I'm not using the inline keyword in some tests. When I let gcc play tricks, a billion calls to pop4() takes 12.56 gigacycles, but I determined it was folding arguments as constant expressions. A more realistic number appears to be 19.6gc for another 30% speed-up. My test loop now looks like this, making sure each argument is different enough to stop gcc from playing tricks.
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();
256 billion bits summed in 8.17s elapsed. Works out to 1.02s for 32 million bits as benchmarked in the 16-bit table lookup. Can't compare directly, because the other bench doesn't give a clock speed, but looks like I've slapped the snot out of the 64KB table edition, which is a tragic use of L1 cache in the first place.
Update: decided to do the obvious and create pop6() by adding four more duplicated lines. Came out to 22.8gc, 384 billion bits summed in 9.5s elapsed. So there's another 20% Now at 800ms for 32 billion bits.
Why not iteratively divide by 2?
count = 0
while n > 0
if (n % 2) == 1
count += 1
n /= 2
I agree that this isn't the fastest, but "best" is somewhat ambiguous. I'd argue though that "best" should have an element of clarity
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.
For a happy medium between a 232 lookup table and iterating through each bit individually:
int bitcount(unsigned int num){
int count = 0;
static int nibblebits[] =
{0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4};
for(; num != 0; num >>= 4)
count += nibblebits[num & 0x0f];
return count;
}
From http://ctips.pbwiki.com/CountBits
This can be done in O(k), where k is the number of bits set.
int NumberOfSetBits(int n)
{
int count = 0;
while (n){
++ count;
n = (n - 1) & n;
}
return count;
}
It's not the fastest or best solution, but I found the same question in my way, and I started to think and think. finally I realized that it can be done like this if you get the problem from mathematical side, and draw a graph, then you find that it's a function which has some periodic part, and then you realize the difference between the periods... so here you go:
unsigned int f(unsigned int x)
{
switch (x) {
case 0:
return 0;
case 1:
return 1;
case 2:
return 1;
case 3:
return 2;
default:
return f(x/4) + f(x%4);
}
}
I think the Brian Kernighan's method will be useful too...
It goes through as many iterations as there are set bits. So if we have a 32-bit word with only the high bit set, then it will only go once through the loop.
int countSetBits(unsigned int n) {
unsigned int n; // count the number of bits set in n
unsigned int c; // c accumulates the total bits set in n
for (c=0;n>0;n=n&(n-1)) c++;
return c;
}
Published in 1988, the C Programming Language 2nd Ed. (by Brian W. Kernighan and Dennis M. Ritchie) mentions this in exercise 2-9. On April 19, 2006 Don Knuth pointed out to me that this method "was first published by Peter Wegner in CACM 3 (1960), 322. (Also discovered independently by Derrick Lehmer and published in 1964 in a book edited by Beckenbach.)"
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.
private int get_bits_set(int v)
{
int c; // 'c' accumulates the total bits set in 'v'
for (c = 0; v>0; c++)
{
v &= v - 1; // Clear the least significant bit set
}
return c;
}
Few open questions:-
If the number is negative then?
If the number is 1024 , then the "iteratively divide by 2" method will iterate 10 times.
we can modify the algo to support the negative number as follows:-
count = 0
while n != 0
if ((n % 2) == 1 || (n % 2) == -1
count += 1
n /= 2
return count
now to overcome the second problem we can write the algo like:-
int bit_count(int num)
{
int count=0;
while(num)
{
num=(num)&(num-1);
count++;
}
return count;
}
for complete reference see :
http://goursaha.freeoda.com/Miscellaneous/IntegerBitCount.html
I use the below code which is more intuitive.
int countSetBits(int n) {
return !n ? 0 : 1 + countSetBits(n & (n-1));
}
Logic : n & (n-1) resets the last set bit of n.
P.S : I know this is not O(1) solution, albeit an interesting solution.
What do you means with "Best algorithm"? The shorted code or the fasted code? Your code look very elegant and it has a constant execution time. The code is also very short.
But if the speed is the major factor and not the code size then I think the follow can be faster:
static final int[] BIT_COUNT = { 0, 1, 1, ... 256 values with a bitsize of a byte ... };
static int bitCountOfByte( int value ){
return BIT_COUNT[ value & 0xFF ];
}
static int bitCountOfInt( int value ){
return bitCountOfByte( value )
+ bitCountOfByte( value >> 8 )
+ bitCountOfByte( value >> 16 )
+ bitCountOfByte( value >> 24 );
}
I think that this will not more faster for a 64 bit value but a 32 bit value can be faster.
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
if you're using C++ another option is to use template metaprogramming:
// recursive template to sum bits in an int
template <int BITS>
int countBits(int val) {
// return the least significant bit plus the result of calling ourselves with
// .. the shifted value
return (val & 0x1) + countBits<BITS-1>(val >> 1);
}
// template specialisation to terminate the recursion when there's only one bit left
template<>
int countBits<1>(int val) {
return val & 0x1;
}
usage would be:
// to count bits in a byte/char (this returns 8)
countBits<8>( 255 )
// another byte (this returns 7)
countBits<8>( 254 )
// counting bits in a word/short (this returns 1)
countBits<16>( 256 )
you could of course further expand this template to use different types (even auto-detecting bit size) but I've kept it simple for clarity.
edit: forgot to mention this is good because it should work in any C++ compiler and it basically just unrolls your loop for you if a constant value is used for the bit count (in other words, I'm pretty sure it's the fastest general method you'll find)
C++20 std::popcount
The following proposal has been merged http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2019/p0553r4.html and should add it to a the <bit> header.
I expect the usage to be like:
#include <bit>
#include <iostream>
int main() {
std::cout << std::popcount(0x55) << std::endl;
}
I'll give it a try when support arrives to GCC, GCC 9.1.0 with g++-9 -std=c++2a still doesn't support it.
The proposal says:
Header: <bit>
namespace std {
// 25.5.6, counting
template<class T>
constexpr int popcount(T x) noexcept;
and:
template<class T>
constexpr int popcount(T x) noexcept;
Constraints: T is an unsigned integer type (3.9.1 [basic.fundamental]).
Returns: The number of 1 bits in the value of x.
std::rotl and std::rotr were also added to do circular bit rotations: Best practices for circular shift (rotate) operations in C++
You can do:
while(n){
n = n & (n-1);
count++;
}
The logic behind this is the bits of n-1 is inverted from rightmost set bit of n.
If n=6, i.e., 110 then 5 is 101 the bits are inverted from rightmost set bit of n.
So if we & these two we will make the rightmost bit 0 in every iteration and always go to the next rightmost set bit. Hence, counting the set bit. The worst time complexity will be O(log n) when every bit is set.
I'm particularly fond of this example from the fortune file:
#define BITCOUNT(x) (((BX_(x)+(BX_(x)>>4)) & 0x0F0F0F0F) % 255)
#define BX_(x) ((x) - (((x)>>1)&0x77777777)
- (((x)>>2)&0x33333333)
- (((x)>>3)&0x11111111))
I like it best because it's so pretty!
Java JDK1.5
Integer.bitCount(n);
where n is the number whose 1's are to be counted.
check also,
Integer.highestOneBit(n);
Integer.lowestOneBit(n);
Integer.numberOfLeadingZeros(n);
Integer.numberOfTrailingZeros(n);
//Beginning with the value 1, rotate left 16 times
n = 1;
for (int i = 0; i < 16; i++) {
n = Integer.rotateLeft(n, 1);
System.out.println(n);
}
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];
}
A fast C# solution using a pre-calculated table of Byte bit counts with branching on the input size.
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
};
}
I always use this in competitive programming, and it's easy to write and is efficient:
#include <bits/stdc++.h>
using namespace std;
int countOnes(int n) {
bitset<32> b(n);
return b.count();
}
There are many algorithm to count the set bits; but i think the best one is the faster one!
You can see the detailed on this page:
Bit Twiddling Hacks
I suggest this one:
Counting bits set in 14, 24, or 32-bit words using 64-bit instructions
unsigned int v; // count the number of bits set in v
unsigned int c; // c accumulates the total bits set in v
// option 1, for at most 14-bit values in v:
c = (v * 0x200040008001ULL & 0x111111111111111ULL) % 0xf;
// option 2, for at most 24-bit values in v:
c = ((v & 0xfff) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
c += (((v & 0xfff000) >> 12) * 0x1001001001001ULL & 0x84210842108421ULL)
% 0x1f;
// option 3, for at most 32-bit values in v:
c = ((v & 0xfff) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
c += (((v & 0xfff000) >> 12) * 0x1001001001001ULL & 0x84210842108421ULL) %
0x1f;
c += ((v >> 24) * 0x1001001001001ULL & 0x84210842108421ULL) % 0x1f;
This method requires a 64-bit CPU with fast modulus division to be efficient. The first option takes only 3 operations; the second option takes 10; and the third option takes 15.
int a = 12;
for eg: binary of 12 is 1100 so answer should be 3 as 3rd bit from right is set.
I want the position of the last most set bit of a. Can anyone tell me how can I do so.
NOTE : I want position only, here I don't want to set or reset the bit. So it is not duplicate of any question on stackoverflow.
This answer Unset the rightmost set bit tells both how to get and unset rightmost set bit for an unsigned integer or signed integer represented as two's complement.
get rightmost set bit,
x & -x
// or
x & (~x + 1)
unset rightmost set bit,
x &= x - 1
// or
x -= x & -x // rhs is rightmost set bit
why it works
x: leading bits 1 all 0
~x: reversed leading bits 0 all 1
~x + 1 or -x: reversed leading bits 1 all 0
x & -x: all 0 1 all 0
eg, let x = 112, and choose 8-bit for simplicity, though the idea is same for all size of integer.
// example for get rightmost set bit
x: 01110000
~x: 10001111
-x or ~x + 1: 10010000
x & -x: 00010000
// example for unset rightmost set bit
x: 01110000
x-1: 01101111
x & (x-1): 01100000
Finding the (0-based) index of the least significant set bit is equivalent to counting how many trailing zeros a given integer has. Depending on your compiler there are builtin functions for this, for example gcc and clang support __builtin_ctz.
For MSVC you would need to implement your own version, this answer to a different question shows a solution making use of MSVC intrinsics.
Given that you are looking for the 1-based index, you simply need to add 1 to ctz's result in order to achieve what you want.
int a = 12;
int least_bit = __builtin_ctz(a) + 1; // least_bit = 3
Note that this operation is undefined if a == 0. Furthermore there exist __builtin_ctzl and __builtin_ctzll which you should use if you are working with long and long long instead of int.
One can use the property of 2s-complement here.
Fastest way to find 2s-complement of a number is to get the rightmost set bit and flip everything to the left of it.
For example: consider a 4 bit system
/* Number in binary */
4 = 0100
/* 2s complement of 4 */
complement = 1100
/* which nothing but */
complement == -4
/* Result */
4 & (-4) = 0100
Notice that there is only one set bit and its at rightmost set bit of 4.
Similarly we can generalise this for n.
n&(-n) will contain only one set bit which is actually at the rightmost set bit position of n.
Since there is only one set bit in n&(-n), it is a power of 2.
So finally we can get the bit position by:
log2(n&(-n))+1
The leftmost bit of n can be obtained using the formulae:
n & ~(n-1)
This works because when you calculate (n-1) .. you are actually making all the zeros till the rightmost bit to 1, and the rightmost bit to 0.
Then you take a NOT of it .. which leaves you with the following:
x= ~(bits from the original number) + (rightmost 1 bit) + trailing zeros
Now, if you do (n & x), you get what you need, as the only bit that is 1 in both n and x is the rightmost bit.
Phewwwww .. :sweat_smile:
http://www.catonmat.net/blog/low-level-bit-hacks-you-absolutely-must-know/
helped me understand this.
There is a neat trick in Knuth 7.1.3 where you multiply by a "magic" number (found by a brute-force search) that maps the first few bits of the number to a unique value for each position of the rightmost bit, and then you can use a small lookup table. Here is an implementation of that trick for 32-bit values, adapted from the nlopt library (MIT/expat licensed).
/* Return position (0, 1, ...) of rightmost (least-significant) one bit in n.
*
* This code uses a 32-bit version of algorithm to find the rightmost
* one bit in Knuth, _The Art of Computer Programming_, volume 4A
* (draft fascicle), section 7.1.3, "Bitwise tricks and
* techniques."
*
* Assumes n has a 1 bit, i.e. n != 0
*
*/
static unsigned rightone32(uint32_t n)
{
const uint32_t a = 0x05f66a47; /* magic number, found by brute force */
static const unsigned decode[32] = { 0, 1, 2, 26, 23, 3, 15, 27, 24, 21, 19, 4, 12, 16, 28, 6, 31, 25, 22, 14, 20, 18, 11, 5, 30, 13, 17, 10, 29, 9, 8, 7 };
n = a * (n & (-n));
return decode[n >> 27];
}
Try this
int set_bit = n ^ (n&(n-1));
Explanation:
As noted in this answer, n&(n-1) unsets the last set bit.
So, if we unset the last set bit and xor it with the number; by the nature of the xor operation, the last set bit will become 1 and the rest of the bits will return 0
1- Subtract 1 form number: (a-1)
2- Take it's negation : ~(a-1)
3- Take 'AND' operation with original number:
int last_set_bit = a & ~(a-1)
The reason behind subtraction is, when you take negation it set its last bit 1, so when take 'AND' it gives last set bit.
Check if a & 1 is 0. If so, shift right by one until it's not zero. The number of times you shift is how many bits from the right is the rightmost bit that is set.
You can find the position of rightmost set bit by doing bitwise xor of n and (n&(n-1) )
int pos = n ^ (n&(n-1));
I inherited this one, with a note that it came from HAKMEM (try it out here). It works on both signed and unsigned integers, logical or arithmetic right shift. It's also pretty efficient.
#include <stdio.h>
int rightmost1(int n) {
int pos, temp;
for (pos = 0, temp = ~n & (n - 1); temp > 0; temp >>= 1, ++pos);
return pos;
}
int main()
{
int pos = rightmost1(16);
printf("%d", pos);
}
You must check all 32 bits starting at index 0 and working your way to the left. If you can bitwise-and your a with a one bit at that position and get a non-zero value back, it means the bit is set.
#include <limits.h>
int last_set_pos(int a) {
for (int i = 0; i < sizeof a * CHAR_BIT; ++i) {
if (a & (0x1 << i)) return i;
}
return -1; // a == 0
}
On typical systems int will be 32 bits, but doing sizeof a * CHAR_BIT will get you the right number of bits in a even if it's a different size
Accourding to dbush's solution, Try this:
int rightMostSet(int a){
if (!a) return -1; //means there isn't any 1-bit
int i=0;
while(a&1==0){
i++;
a>>1;
}
return i;
}
return log2(((num-1)^num)+1);
explanation with example: 12 - 1100
num-1 = 11 = 1011
num^ (num-1) = 12^11 = 7 (111)
num^ (num-1))+1 = 8 (1000)
log2(1000) = 3 (answer).
x & ~(x-1) isolates the lowest bit that is one.
int main(int argc, char **argv)
{
int setbit;
unsigned long d;
unsigned long n1;
unsigned long n = 0xFFF7;
double nlog2 = log(2);
while(n)
{
n1 = (unsigned long)n & (unsigned long)(n -1);
d = n - n1;
n = n1;
setbit = log(d) / nlog2;
printf("Set bit: %d\n", setbit);
}
return 0;
}
And the result is as below.
Set bit: 0
Set bit: 1
Set bit: 2
Set bit: 4
Set bit: 5
Set bit: 6
Set bit: 7
Set bit: 8
Set bit: 9
Set bit: 10
Set bit: 11
Set bit: 12
Set bit: 13
Set bit: 14
Set bit: 15
Let x be your integer input.
Bitwise AND by 1.
If it's even ie 0, 0&1 returns you 0.
If it's odd ie 1, 1&1 returns you 1.
if ( (x & 1) == 0) )
{
std::cout << "The rightmost bit is 0 ie even \n";
}
else
{
std::cout<< "The rightmost bit is 1 ie odd \n";
}```
Alright, so number systems is just working with logarithms and exponents. So I'll dive down into an approach that really makes sense to me.
I would prefer you read this because I write there about how I interpret logarithms as.
When you perform the x & -x operation, it gives you the value which has the right most bit as 1 (for example, it can be 0001000 or 0000010. Now according to how I interpret logarithms as, this value of the right most set bit, is the final value after I grow at the rate of 2. Now we are interested in finding the number of digits in this answer because whatever that is, if you subtract 1 from it, that is precisely the bit-count of set bit (bit count begins with 0 here and the digit count begins with 1, so yeah). But the number of digits is precisely the time you expanded for + 1 (in accordance with my logic) or just the formula I mentioned in the previous link. But now, as we don't really need the digits, but need the bit count, and we also don't have to worry about values of bits which potentially can be real (if the number is 65) because the number is always some multiple of 2 (except 1). So if you just take the logarithm of the value x & -x, we get the bit count! I did see an answer before that mentioned this, but diving down to why it really works was something I felt like writing down.
P.S: You could also count the number of digits and then subtract 1 from it to get the bit-count.
This question already has answers here:
What is the fastest/most efficient way to find the highest set bit (msb) in an integer in C?
(31 answers)
Closed 8 years ago.
Just like the title said, I have a char (e.g. 00101011), how can I know the position of first '1' bit in it? (in this case, it is 6, or 5 if we consider the rightmost position is 0).
I know a for-loop shift would be helpful, but is there a more efficient way?
There are some bithacks (as mentioned by Codor) which can do it better than just shifting, but nothing in O(1). Binary search comes to mind.
The only more efficient way I am afraid is a specialised processor instruction like the bsr instruction on x86. Some compilers provide builtins for those instructions. gcc for example provides __builtin_clz for this purpose.
Various approaches for getting the highest 1 bit are discussed here, where also the shifting method you described is mentioned. A detailed answer from there:
Find the log base 2 of an integer with a lookup table
static const char LogTable256[256] =
{
#define LT(n) n, n, n, n, n, n, n, n, n, n, n, n, n, n, n, n
-1, 0, 1, 1, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
LT(4), LT(5), LT(5), LT(6), LT(6), LT(6), LT(6),
LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7), LT(7)
};
unsigned int v; // 32-bit word to find the log of
unsigned r; // r will be lg(v)
register unsigned int t, tt; // temporaries
if (tt = v >> 16)
{
r = (t = tt >> 8) ? 24 + LogTable256[t] : 16 + LogTable256[tt];
}
else
{
r = (t = v >> 8) ? 8 + LogTable256[t] : LogTable256[v];
}
The lookup table method takes only about 7 operations to find the log of a 32-bit value. If extended for 64-bit quantities, it would take roughly 9 operations. Another operation can be trimmed off by using four tables, with the possible additions incorporated into each. Using int table elements may be faster, depending on your architecture.
The code above is tuned to uniformly distributed output values. If your inputs are evenly distributed across all 32-bit values, then consider using the following:
if (tt = v >> 24)
{
r = 24 + LogTable256[tt];
}
else if (tt = v >> 16)
{
r = 16 + LogTable256[tt];
}
else if (tt = v >> 8)
{
r = 8 + LogTable256[tt];
}
else
{
r = LogTable256[v];
}
To initially generate the log table algorithmically:
LogTable256[0] = LogTable256[1] = 0;
for (int i = 2; i < 256; i++)
{
LogTable256[i] = 1 + LogTable256[i / 2];
}
LogTable256[0] = -1; // if you want log(0) to return -1
Behdad Esfahbod and I shaved off a fraction of an operation (on average) on May 18, 2005. Yet another fraction of an operation was removed on November 14, 2006 by Emanuel Hoogeveen. The variation that is tuned to evenly distributed input values was suggested by David A. Butterfield on September 19, 2008. Venkat Reddy told me on January 5, 2009 that log(0) should return -1 to indicate an error, so I changed the first entry in the table to that.
First of all, bits in a byte are always enumerated from 0 to 7, where bit 0 is the least significant bit. This way of counting bits is universal.
As for what's most efficient, define "efficient". The most speed-efficient way is to generate a table of 256 bytes at compile time, where each item contains the bit position of the first bit:
#define ALL_ZEROES 0xFF // special case
const uint8_t LOOKUP_TABLE [256] =
{
ALL_ZEROES, // no 1 found
0, // 0000 0001, first 1 at bit 0
1, // 0000 0010, first 1 at bit 1
0, // 0000 0011, first 1 at bit 0
2, // 0000 0100, first 1 at bit 2
...
};
And then you get the position by typing:
first_one = LOOKUP_TABLE[my_value];
However, this does consume 256 bytes of ROM memory, just to gain a few extra CPU ticks. So while it is very speed-efficient, it is horribly inefficient memory-wise.
The mentioned method with a for loop and shift is already fast, and it doesn't come with this memory overhead. So in most applications you would probably use the for loop, because it is most efficient in terms of speed + memory consumption together.
It all depends on the nature of the application.