FreeLists / openbeos / [openbeos] Re: app_server: MMX/SSE help wanted

[Adi Oanca <adioanca@xxxxxxxxxxxxxx>]

OK, there seem to be some problems with AMD CPUs (maybe.) It doesn't matter: we will write the same 
code for both. Where we will have problems we will implement for each vendor.

Now, I have another question(s):
	If GCC 3, app_server, media_kit and translators will all use MMX, how can we be sure they won't 
collide with each other?
	MMX runs in parallel with the CPU, but which thread has access to MMX registers? The kernel also 
manages this? How can MMX be used, do we need a separate thread with MMX loops only?

	What about the case where app_server_MMX is compiled with GCC3_MMX ? Should GCC MMX support be 
disabled in this case?


Thanks,
Adi.


Christian Packmann wrote:
> On 2004-08-08 17:29:04 [+0000], Kian Duffy wrote:
>
> > On Sun, 08 Aug 2004 19:22:10 +0000, Christian Packmann 
> > <christian.packmann@xxxxxx> wrote:
> >
> > > On 2004-08-08 17:00:18 [+0000], Kian Duffy wrote:
>
>  
>
> > > > Well, if it was a true clone of Intel's SSE, it wouldn't crash BeOS. 
> > > > It is a bad clone of it, hence the crash
>
>
> > > Okay, then let me know the instruction sequence on which an XP crashes, 
> > > but a PIII/PIV doesn't. I've already checked the Intel docs 
> > > (admittedly, not page to page), and so far I haven't found anything 
> > > which would indicate an incompatibility.
>
>  
>
> > I'd suggest you find someone with an Athlon XP to ask, I no longer posess 
> > any AMD kit due
>
>
> I do have an Athlon XP, it just hasn't misbehaved so far in regard to SSE 
> code. That's why I'd be interested in getting such instruction sequences. 
>
> I've heard the myth about the 'incompatible' Athlon XP dozens of times, but 
> never an explanation which I could verify empirically. I'm just a little 
> tired of that bit of misinformation floating around, that's all.
>
>
> > to the number of times my Duron went on fire (at rated 
> > speed with above rated cooling).
>
>
> Probably wrong cooler or mounting error. I once had a mainboard with 
> premounted Duron+cooler, and that also ran very hot, 60-70°. Now I've got a 
> XP2100+ and good cooler, the XP reaches ca. 55° under full load in summer; 
> in winter it's overclocked without problems.
>
>
> > > > And hence the problems with the MMX optomised version of the Windows
> > > > 95 game "POD" on AMD systems.
>
>
> > > What kind of problems, especially problems relating to the XP?
>
>
> > The problems I had were not on an XP (I no longer use AMD kit, see 
> > above), but on a Spitfire Duron. Basically, the game GPF'ed on loading 
> > due to a misintrpretation of a MMX instruction. As I no longer have an 
> > AMD system (or a Windows system), I can't check exactly what it was.
> >
> > You needed the specific MMX version that came with Dell Dimension 
> > H-series systems for that behaviour to happen, older versions were not 
> > MMX enabled and later MMX versions worked on AMD CPU's with MMX.
>
>
> This sounds more like a program error than h/w error. I've never heard of a 
> problem in the MMX engine of the K7 core. And searching Google for "POD MMX 
> AMD problem" just turned up hits on problems with the Forcefeedback driver, 
> which are solved by patching the driver.
>
> If the MMX core really was buggy you'd hear a lot more complaints, as all 
> software using that specific MMX command would be crashing. And MMX is 
> being used pretty widely now, not only for graphics/sound processing but 
> for prime search (Prime95, Seventeen or Bust), probably SETI, RC5, OGR and 
> others as well.
>
> Bye,
> Chris