libpopcnt

libpopcnt.h is a header-only C/C++ library for counting the number of 1 bits (bit population count) in an array as quickly as possible using specialized CPU instructions i.e. POPCNT, AVX2, AVX512, NEON, SVE. libpopcnt.h has been tested successfully using the GCC, Clang and MSVC compilers.

C/C++ API

#include "libpopcnt.h"

/*
 * Count the number of 1 bits in the data array
 * @data: An array
 * @size: Size of data in bytes
 */
uint64_t popcnt(const void* data, uint64_t size);

How to compile

libpopcnt.h does not require any special compiler flags like -mavx2! To get the best performance we only recommend to compile with optimizations enabled e.g. -O3 or -O2.

cc  -O3 program.c
c++ -O3 program.cpp

How it works

On x86 CPUs, libpopcnt.h first queries your CPU's supported instruction sets using the CPUID instruction (this is done only once). Then libpopcnt.h chooses the fastest bit population count algorithm supported by your CPU:

• If the CPU supports AVX512 the AVX512 VPOPCNT algorithm is used.
• Else if the CPU supports AVX2 the AVX2 Harley Seal algorithm is used.
• Else if the CPU supports POPCNT the POPCNT algorithm is used.
• For CPUs without POPCNT instruction a portable integer algorithm is used.

Note that libpopcnt.h works on all CPUs (x86, ARM, PPC, WebAssembly, ...). It is portable by default and hardware acceleration is only enabled if the CPU supports it. libpopcnt.h it is also thread-safe.

We take performance seriously, if you compile using e.g. -march=native on an x86 CPU with AVX512 support then all runtime CPUID checks are removed!

CPU architectures

libpopcnt.h has hardware accelerated popcount algorithms for the following CPU architectures:

For other CPU architectures a fast integer popcount algorithm is used.

Development

cmake .
make -j
make test

The above commands also build the benchmark program which is useful for benchmarking libpopcnt.h. Below is a usage example run on an Intel Xeon Platinum 8168 CPU from 2017:

# Usage: ./benchmark [array bytes] [iters]
./benchmark
Iters: 10000000
Array size: 16.00 KB
Algorithm: AVX512
Status: 100%
Seconds: 1.59
103.4 GB/s

Acknowledgments

Some of the algorithms used in libpopcnt.h are described in the paper Faster Population Counts using AVX2 Instructions by Daniel Lemire, Nathan Kurz and Wojciech Mula (23 Nov 2016). The AVX2 Harley Seal popcount algorithm used in libpopcnt.h has been copied from Wojciech Muła's sse-popcount GitHub repo.