Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
    Zachariah Etienne
    @zachetienne

    Has anyone else used SLEEF as a fast SIMD library http://sleef.org/ ? Are there any nice open-source alternatives out there?

    Also, in case you didn't catch the news, the latest glibc (2.27) contains new, faster vectorized math functions (contributed by Intel) https://www.phoronix.com/scan.php?page=news_item&px=Intel-AVX2-FMA-Math-Glibc-2.27 ; https://lwn.net/Articles/746327/ 50% speed-ups on Skylake for some functions are reported.

    Zachariah Etienne
    @zachetienne
    Is ETK capable of handling SIMD-vectorized transcendental functions with Kranc-generated code? If so, how is it done?
    Roland Haas
    @rhaas80
    I would look at vecmathlib: https://bitbucket.org/eschnett/vecmathlib which is however not what Kranc does since Kranc uses the older Vectors thorn.
    Which also has vectorized versions in its kcos etc functions.
    Eg https://bitbucket.org/cactuscode/cactusutils/src/500fda1ed409a5435aa8e121c8b4802c035b4955/Vectors/src/vectors-4-AVX.h#lines-463
    Erik Schnetter
    @eschnett
    Zach -- This is handled by thorn Vectors. In short, if the CPU offers respective instructions, then we use them, otherwise the functions are scalarized (which is very expensive).
    As Roland says, the vecmathlib library has implementations (fast, not very accurate, and they don't handle all IEEE cases correctly) that are vectorized.
    there is also SLEEF http://sleef.org. this is highly optimized, probably better than vecmathlib, but does not have a convenient C++ interface.
    Zachariah Etienne
    @zachetienne
    @rhaas80 , @eschnett : Thanks, this is very useful. I'll take a look at Vectors & vecmathlib.
    Roland Haas
    @rhaas80
    There is some restriction possible. We have them enabled for cactuscode/cactus: https://bitbucket.org/cactuscode/cactus/admin/branch-permissions
    sorry, wrong chat