[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Bug#909457: ITP: blis -- BLAS-like Library Instantiation Software Framework



It makes sense to set BLIS's priority to 37, because

1. BLIS doesn't enable threading by default, but
   BLIS is faster than (default) Atlas in single-thread mode.

   Hence BLIS > Atlas

2. BLIS has comparative performance compared to OpenBLAS/MKL, if we
   compile it with openmp threading and use it with a proper
   BLIS_NUM_THREADS configuration. My test results suggest that
   BLIS is sometimes slightly faster than OpenBLAS and sometimes
   slower. Anyway if we don't touch the environment variable,
   openblas is always faster than BLIS.

   Hence OpenBLAS > BLIS

I posted my simple benchmark result here:
https://github.com/flame/blis/issues/255


On Mon, Sep 24, 2018 at 07:06:44AM +0000, Mo Zhou wrote:
> Package: wnpp
> Severity: wishlist
> X-Debbugs-CC: debian-science@lists.debian.org
> Owner: Mo Zhou <lumin@debian.org>
> 
> * Package name    : blis
>   Version         : 0.4.1
>   Upstream Author : The University of Texas at Austin, HP Enterprise, AMD Inc.
> * URL             : https://github.com/flame/blis
> * License         : BSD-3-clause
>   Programming Lang: C
>   Description     : BLAS-like Library Instantiation Software Framework
> 
> This is a relatively new BLAS implementation and a new candidate to
> libblas.so.3 . However unlike OpenBLAS, BLIS doesn't provide LAPACK API/ABI.
> 
> Upstream benchmarks shows a promising and comparative performance compared
> to OpenBLAS and MKL, however I'm still not quite sure how fast it is,
> objectively. So maybe I'll set the priority to 37, which means for the
> libblas.so.3 alternative,
> 
>   OpenBLAS > BLIS > Atlas > Netlib > MKL
> 
> Moreinfo: https://github.com/flame/blis/issues/254
> 


Reply to: