Re: [AD] Use MMX to get fast

[aj <aj@xxxxxxxxxx>]

At 06:59 AM 30/09/2004, you wrote:
> > >It's basically replacing a normal memcpy with an MMX QWORD copy,
> >
> > how do you know memcpy doesn't do that already ?
>
> Well ok, the 'C' library memcpy might, but what I meant was allegro
> currently uses the standard assembly language copy (rep stosl, etc).

then your mmx qword copy  would be better yes?
if so, please can you try to implement it for allegro.
i am always keen to see faster blit code.


> Anyway yes, any improvements optimised especially for mem->mem blits would
> definitely be a good thing.


sorry if i seem against it, im not, at all.. im all for it..
if the current allegro 'rep stosl' copy, is slow (your opinion welcome 
here),  and a memcpy does contain mmx-qword (faster) copying.. would 
allegro benefit from using a memcpy  instead of the 'rep stosl' copy ?

what are the limitations of using this mmx-qword copy ?
i guess 8byte aligned,  cpu has mmx available,  what else ?
is the cost of setting up the mmx-qword copy high?   will it be ok for 
small bitmaps ?
would calling memcpy() have any overhead?   i guess its speed is 
implementation/platform specific.


aj.