> On 16 Jun, linker@nightshade.ml.org wrote:
> > Wouldn't it be better for all if apps:
> >
> > i = detect_MMMX();
> > if !i then usenormalstuff=1;
> > else usemmx=1;
> >
> > then had code for both cases???
>
> no....
>
> If you can assume the MMX functions are there then EACH application
> doesn't have to emulate all of the MMX stuff itself if it doesn't have
> the REAL hardware.
>
> This should be treated EXACTLY like the Floating Point hardware... at
> least in my opinion.
Hmm.. The whole reason for using MMX is so that you can get a great speed
improvement on a VERY limited class of functions. Usually you end up hand
coding the new MMX function.
Using MMX emu will completely kill performance, it makes the mmx opts
absolutly pointless.
The FPU is useful to an incredbily wide range of formulas.. Infact, if you
dont have FPU or a FPU emulator there are quite a few things you can not
do (without basicly emulating the FPU in software in the app itself).
Everything that MMX can do can be done with standard int code, just
slower.
I would dare to say that you almost always have a clean C implimentation
of your stuff that you want to MMXify before you actually MMXify it.
Furthermore, the idea of a FPU exists on all platforms Linux runs on
(no?).
MMX is a cpu specific optimazation.. You suggestion would is like:
We need to include a x86 opcode emulator in the kernel, so that Alphas can
execute apps where there programmer wrote the program in C but a few
functions in x86 asm and didn't provide C versions. Or vice versa (alpha
code on x86).
That case would hopefully never occure.. A MMX only app wouldn't run on
PPC/Alpha/Mips/68k/Stronarm/non-mmx pentium.
If the programmer wants portability he will have to provide a C version
along with the optimized version(s). It's not hard.
> -jeffrey hundsatd
Gregory Maxwell
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu