Re: LuaJIT2 performance for number crunching

[KHMan <keinhong@...>]

On 2/16/2011 5:38 PM, Miles Bader wrote:
> KHMan writes:
> > > rk4-unroll.lua with LuaJIT2: 0m4.391
> > > C code with GSL lib (opt -O2): 0m20.094
> >
> > Sorry to barge in, it is a worrying difference. One thing is bugging me:
> > Is the C code running SSE2? IIRC gcc -O2 does not normally enable SSE2.
>
> It will enable it by default when it thinks it's possible.

But this is not standard behaviour. I haven't heard of such a 
change in gcc default behaviour from the past few release notices. 
Would be much obliged if you can correct me on this.

Checked the gcc 4.5.2 docs, AFAICT there is still a generic 
behaviour and it's not "-march=native". Still need to explicitly 
enable SSE2 if the compiler was not built with --with-fpmath=sse 
or something like that.

> It's always a good idea to use "-march=native" if you're looking for the
> best performance...

And those "software automatic tuning" guys from the 
high-performance computing side will say that there is a lot more 
to do in order to wring the last bit of speed from a C compiler 
with a ton of options... :-)

--
Cheers,
Kein-Hong Man (esq.)
Kuala Lumpur, Malaysia