Re: [ublas] Matrix multiplication performance

Subject: Re: [ublas] Matrix multiplication performance
From: Michael Lehn (michael.lehn_at_[hidden])
Date: 2016-01-24 11:21:10

On 24 Jan 2016, at 17:02, Peter Schmitteckert <Peter.Schmitteckert_at_[hidden]> wrote:

> Dear All,
> 
> 
> first of all I'm impressed by the recent results.
> 
> > We need both native implementations and bindings.
> 
> I completely agree.

I also agree.  But that is no contradiction of specifying a unified low-level C++BLAS standard.  One
possible and obvious implementation would be a binding to BLAS or CBLAS.  Also there is no real
BLAS or CBLAS standard like in “ANSI” or “ISO”.  Its just a de facto standard.  So as long as it is
useful it will become a de facto standard.  Even if it just is the nicest way to be a BLAS-binding it
should be attractive for other C++ libraries like MTL4 or BLAZE.

Its also amazing that libraries like ATLAS, OpenBLAS or BLIS are written in C and not in C++.  I am
pretty familiar with their code and using C++ would make their implementation so much easier. And
I am not talking about OOP.  Just using template functions and doing the rest in C-style.  Basically
they use macros for generating functions like sgemm, dgemm, cgemm, zgemm from a single
implementation.  Very unreadable, believe me.  So even for a BLAS implementation that just has
to support a C and Fortran interface my choice would be C++ for the actual implementation.  And
I am certainly not C++ fan boy.  Really.