[ODE] Intel/Msvc Compiler
Tom De Muer
tom.demuer at skynet.be
Tue Jan 6 10:05:00 MST 2004
Hi,
> Hi Tom,
>
> TDM> Just did a little test yesterday, I compiled ODE single precision
(the msvc7
> TDM> contrib/default) with MSVC and got my car simulated at +/- 700 fps,
the
> TDM> intel compiled code run at 870 fps.
>
> I'd bet the 700 and 870 fps stuff doesn't say much thing about
> real performance. It just indicates that your computer is too
> powerful compare to your program.
>
> I only believe in profilers. They show much more "real" info
> than just a handmade FPS counter.
>
Well the FPS I'm stating is the theoretically maximum of the _physics_
nothing else, thus in other words it says that with the same code, the msvc
compiler produces an ode.dll which makes the physics part of my sim running
at an average of 1/700 s and with the intel at 1/870 s. The global
framerate of my "thing" is a lot lower because there are other things that
need be done. I agree that the numbers quoted may not have large precision
but the difference is significant enough.
> TDM> The intel compiler also has automatic
> TDM> vectorisation that converts regular loops on floats into loops with
SIMD but
> TDM> my Athlon only has limited simd support and it still is the _intel_
compiler
> TDM> ;)... I thought it may be interesting in the people trying
optimising stuff
> TDM> as I see topics about that reguraly popping up :o) Has anybody else
tried
> TDM> to fiddle with the intel/msvc compiler settings for optimal
performance?
>
> I didn't play much with them but IMHO, the quality of the code is
> more important than what compiler to use.
>
I agree fully that algorithmic optimisation is the first step. What I'm
trying to say is that it should also be your last in most cases because the
compiler does the lowlevel optimisation better than you can... that was the
point I'm trying to make and my question also, has anyone experimented with
this stuff too?
> --
> Best regards,
>
Cheers,
Tom
> ---------------------------------------------------------------------
> Nguyen Binh
More information about the ODE
mailing list