Releases · kimwalisch/libpopcnt

29 Jun 09:21

5214d3f

This release improves the AVX512 popcount algorithm. libpopcnt-3.1 is fully backwards compatible with libpopcnt-3.0 and libpopcnt-2.*.

Improve AVX512 popcount algorithm for trailing 64 bytes.
AVX512 algorithm does not require AVX512-BITALG extension anymore.

Assets 2

27 Jun 18:39

kimwalisch

v3.0

972333a

libpopcnt-3.0

libpopcnt-3.0 is a major new release with many improvements, but it is still backwards compatible with libpopcnt-2.*!

The two main new features of libpopcnt-3.0 are: the new ARM SVE popcount algorithm that is up to 3x faster than the ARM NEON popcount algorithm and the new AVX512 VPOPCNT algorithm that is up to 35% faster than the old AVX512 Harley-Seal popcount algorithm. Unlike the old AVX512 algorithm, the new AVX512 VPOPCNT algorithm is also fast for short arrays ≥ 48 bytes.

Add ARM SVE algorithm.
Replace AVX512BW algorithm by faster AVX512 VPOPCNTDQ algorithm.
Add MSVC support for ARM NEON.
Improve preprocessor checks using __has_include() macro.
Port tests from AppVeyor to GitHub actions.
Get rid of unaligned uint64_t memory acceses, this fixes test failures when using GCC compiler sanitizers.
Prefix all libpopcnt macros using LIBPOPCNT_ to avoid any naming collisions.

Assets 2

17 Mar 09:07

kimwalisch

v2.6

2ca88fd

libpopcnt-2.6

What's Changed

Fix building on NetBSD by @brad0 in #18

New Contributors

@brad0 made their first contribution in #18

Full Changelog: v2.5...v2.6

Contributors

brad0

Assets 2

11 Aug 09:58

kimwalisch

v2.5

c49987e

libpopcnt-2.5

On x86/x64 runtime CPUID checks are now removed if the user compiles his code with e.g. -march=native (or -mavx512bw).
On CPU architectures that support unaligned memory accesses, stop aligning memory as it causes branch mispredictions which significantly deteriorate performance for small array sizes.

Assets 2