Berenger Bramas - Inria - berenger.bramas@inria.fr

AVX-512 sort functions

This project is a small library that provides fast functions to sort arrays of int or double or int[2] using AVX-512. A paper describes the different strategies and has been published in International Journal of Advanced Computer Science and Applications(IJACSA), Volume 8 Issue 10, 2017 http://thesai.org/Publications/ViewPaper?Volume=8&Issue=10&Code=IJACSA&SerialNo=44 and a draft is also available here https://arxiv.org/abs/1704.08579 for the KNL (older report). If you use this software, we would appreciate that you cite the related IJACSA paper.

The branch paper contains some not very clean files that were used for benchmarks, wherease the current master branch provides an header only library:

• sort512.hpp : the library that can be directly included in any code to sort integer or double
• sort512kv.hpp : the library that can be directly included in any code to sort key/value pairs of integers
• sort512test.cpp : some unit tests (can be used for examples)

Note that the official repository is https://gitlab.inria.fr/bramas/avx-512-sort

There is a version of this algorithm for ARM SVE here https://gitlab.inria.fr/bramas/arm-sve-sort

Functions

• Sort512::Sort(); to sort an array
• Sort512::SortOmp(); to sort in parallel (need openmp)
• Sort512::Partition512(); to partition
• Sort512::SmallSort16V(); to sort a small array (should be less than 16 AVX512 vectors)

AVX 512 compilation flags (KNL)

• Gcc : -mavx512f -mavx512pf -mavx512er -mavx512cd
• Intel : -xCOMMON-AVX512 -xMIC-AVX512

AVX 512 compilation flags (SKL)

• Gcc : -mavx512f -mavx512cd -mavx512vl -mavx512bw -mavx512dq
• Intel : -xCOMMON-AVX512 -xCORE-AVX512

ARCH compilation flags

Use "-march=native -mtune=native" if you are already on the right platform ("native can be replaced by "knl" or "skylake")

OpenMP compilation flags

In case you want to use the parallel sort, you need to add the flag:

• Gcc : -fopenmp
• Intel : -qopenmp

Using Intel SDE

Anyone can test the code without having a KNL by using the Intel SDE

Knowing more about the number of instructions

Please checkout the branch "feature/counters" to have access to counters.