Re: [Gomp-discuss] fortran c/c++

[Biagio Lucini]

On Wed, 18 Dec 2002, Lars Segerlund wrote:

> 
>   Well the thing is, the 3DNow stuff in the athlons can do 32 bit 
> floatingpoint this way :-), so I think it might be worth considering 
> since the only thing I have seen able to keep up with these numbers ate 
> alphas, but for now I'm with you.
> 
>   / Lars Segerlund.
> 

gcc has a good support for extended instruction sets anyway. I have tried
to see what it can do compared with the Intel compiler, and I can't see
much difference on a Pentium IV (which, by the way, with SSE2 can do also 
64 bit real operations). Of course, if you need specific optimisation, you
will have to write your assembly or in-line assemby by yourself, and the
only way SWAR can help is by providing an interface to make it portable.
But a SWAR-like project looks really complicated to me (and maybe also to
them, given that they appear not to be working on it anymore). It is not
easy to provide general paradigms for automatically optimising code on an
architecture. which is what SWAR is meant for. If you want to be sure of
the optimisation achieved, then you are left with low level programming.
I've paied attention myself to this problem for SSE2 instructions, and in
the end I have decided that what people are doing is good enough for me:
I'm looking for portability, and I don't want to be tied to an
architecture. I don't want either to have to write a new SWAR compiler if
I have to change architecture :-)

Cheers
Biagio

P.S.: About claims about performance, I've heard that around 1GFlops can
be obtained by coding directly SSE2 instructions for critical parts of the
code.