Re: [hatari-devel] Falcon speed

[Thomas Bergstrom <tb@xxxxxxxxxxxxx>]

I use the simple counter method to check CPU consumption.
No DSP, pure 030 with 90-95% running in i-cache or so I guess.

No, I havn't used the profiler, perhaps it is time. :)

Br,

/Thomas



> How do you measure the CPU consumption on 030?
>
> And is DSP running?
>
> (You could check whether you see same difference with "--dsp none")
>
>
> > which makes it a bit hard to know how much work can be put on the CPU
> > and still keep the frame rate on a real Falcon.
>
> Have you used Hatari profiler?  That shows how many emulated
> cycles Hatari spends on each instruction in your code:
> http://hg.tuxfamily.org/mercurialroot/hatari/hatari/raw-
> file/tip/doc/manual.html#Profiling
>
> If you see in that output something suspicious, it would be
> great if you could provide a more isolated use-case demonstrating that.
>
>
> 	- Eero
>
> > I was curious to hear other peoples experiences regarding this. I use
> > Hatari alot and think it's fantastic, just have to find a way to deal
> > with this "problem".
> ...
> > > > I started a game at the same level without playing anything, so these
> > > > are game "idle" CPU Load numbers:
> > > >
> > > > Machine			CPU Load
> > > > --------------------------------
> > > > Falcon			58.05%
> > > > Hatari, CE&PFM Disabled 41.61%
> > > > Hatari, CE&PFM Enabled	76.51%
> >
> > > Are you measuring Hatari's CPU usage, or running game under MiNT
> > > in Hatari and using e.g. "top" to see your game's CPU usage under MiNT?
> > >
> > > If you're measuring Hatari's CPU usage, more accurate emulation
> > > needing more CPU "works as expected".
> > >
> > > 	- Eero
> > >
> > > > Tried on both laptops, both give the exact same numbers. As you can
> > > > see Cycle Exact and Prefetch Mode slows Hatari down to be quite alot
> > > > slower than the real thing. The Hatari version I was running was
> > > > built Jan 2:nd.
> > > >
> > > > I run the windows version, maybe the difference is in Windows vs.
> > > > Linux?
> > > >
> > > > Perhaps it's the data cache but would that really have that much
> > > > impact?
> > > >
> > > > Get the feeling that there is some option/setup that's missing or so.
> > > >
> > > > Br,
> > > >
> > > > /Thomas
> > > >
> > > > On Wed, 7 May 2014, Eero Tamminen wrote:
> > > > > Hi,
> > > > >
> > > > > Hatari WinUAE CPU core doesn't emulate 030 data cache.
> > > > >
> > > > > According to comments by Douglas about his experience
> > > > > on optimization things related to that in Bad Mood, code
> > > > > using data cache well works faster on real Falcon than
> > > > > in Hatari.
> > > > >
> > > > > 	- Eero
> > > > >
> > > > > On keskiviikko 07 toukokuu 2014, laurent.sallafranque@xxxxxxx wrote:
> > > > > > > Disabling them results in a much faster Falcon. Strange that we get
> > > > > > > so different results
> > > > > >
> > > > > > If you say that it's strange that disabling cycle exact and prefetch
> > > > > > mode does much different results and when they're enabled, that's
> > > > > > normal. The table of cycles for the new cpu core is applied only for
> > > > > > this configuration.
> > > > > >
> > > > > > In non cycle exact mode, or non prefetch mode or in MMU mode, each
> > > > > > instruction takes 2 cycles (which is wrong of course)
> > > > > >
> > > > > >
> > > > > > If yo say that what is strange is hatari speed on your machine and
> > > > > > on mine, I agree. I've got a demo that seems to run slower on
> > > > > > hatari than on my real falcon (underscore demo if I'm right) (it
> > > > > > starts in a tunnel, then, there's a 3d object with motion blur,
> > > > > > then, a kind of maze in 3D DSP (that crashs under hatari at the
> > > > > > moment).
> > > > > >
> > > > > > Except this demo, I always detect that hatari runs faster than my
> > > > > > falcon (but I'm running my falcon in VGA, this may change my
> > > > > > perception of speed).
> > > > > >
> > > > > > Regards
> > > > > >
> > > > > > Laurent
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > ----- Mail original -----
> > > > > > De: "Thomas Bergstrom" <tb@xxxxxxxxxxxxx>
> > > > > > À: hatari-devel@xxxxxxxxxxxxxxxxxxx
> > > > > > Envoyé: Mercredi 7 Mai 2014 09:51:55
> > > > > > Objet: Re: [hatari-devel] Falcon speed
> > > > > >
> > > > > >
> > > > > > Hi,
> > > > > >
> > > > > > Hmm. When enabling both cycle exact and prefetch mode I think Hatari
> > > > > > is close to real Falcon speed. That goes for both my laptops, one
> > > > > > AMD A4 and one Intel i5. Disabling them results in a much faster
> > > > > > Falcon. Strange that we get so different results. I have the same
> > > > > > configuration as you except that I have the blitter emu disabled.
> > > > > >
> > > > > > Perhaps my memory is misleading me, I will do some more tests and
> > > > > > get back. :)
> > > > > >
> > > > > > Br,
> > > > > >
> > > > > > /Thomas
> > > > > >
> > > > > >
> > > > > >
> > > > > > On Wed, 7 May 2014, laurent.sallafranque@xxxxxxx wrote:d
> > > > > >
> > > > > > > Hi Thomas,
> > > > > > >
> > > > > > > The Falcon emulation is not speed accurate actually, neither cycle
> > > > > > > accurate.
> > > > > > >
> > > > > > > For example, the current dev version of Racer runs in 3 VBLs under
> > > > > > > Hatari, whereas it runs in 5 VBLs under my Falcon. (Hatari is 2
> > > > > > > VBLs faster than the real machine).
> > > > > > >
> > > > > > > On my fast computer, hatari is always faster than my real Falcon
> > > > > > > (anyway I choose cycle exact, prefetch mode, or not).
> > > > > > >
> > > > > > > The Hatari parameters I use for Falcon emulation are :
> > > > > > >
> > > > > > > CPU : winUAE core (hatari_falcon exe)
> > > > > > >
> > > > > > > CPU Type   : 68030
> > > > > > > CPU clock  : 16 Mhz
> > > > > > > Falcon DSP : Full
> > > > > > >
> > > > > > > CPU parameters :
> > > > > > > 24 bits addressing    : ON
> > > > > > > Prefetch mode, slower : ON
> > > > > > > Cycle exact, slower   : ON
> > > > > > > Real time clock emu   : ON
> > > > > > > Patch Timer_D         : OFF
> > > > > > > Blitter emu           : ON
> > > > > > > MMU emulation         : OFF (except if you need special apps, but
> > > > > > > MMU emulation is not cycle accurate at all)
> > > > > > >
> > > > > > > FPU : 68882
> > > > > > > More compatible, slower : ON
> > > > > > >
> > > > > > >
> > > > > > > TOS : 4.04 Fr
> > > > > > >
> > > > > > > Memory : 4 or 14 megs
> > > > > > >
> > > > > > >
> > > > > > > The cycle exact mode is the one that allows the most programs to
> > > > > > > work (but if you use the old CPU core (hatari.exe) in Falcon mode,
> > > > > > > the timings are again different and some rare programs that don't
> > > > > > > work with the ne core works with the ols one (like the MOAI96 demo
> > > > > > > by Mikro)
> > > > > > >
> > > > > > > The MMU emulation CPU interprets all the CPU instructions (MMU,
> > > > > > > CPU, FPU) in 2 cycles, no matter the instruction.
> > > > > > >
> > > > > > >
> > > > > > > There's still work to do before hatari becomes an accurate Falcon
> > > > > > > emulator, but it's already useable and quite efficient with many
> > > > > > > programs ans demos.
> > > > > > >
> > > > > > > Regards
> > > > > > >
> > > > > > > Laurent
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > ----- Mail original -----
> > > > > > > De: "Thomas Bergstrom" <tb@xxxxxxxxxxxxx>
> > > > > > > À: hatari-devel@xxxxxxxxxxxxxxxxxxx
> > > > > > > Envoyé: Mercredi 7 Mai 2014 09:05:23
> > > > > > > Objet: [hatari-devel] Falcon speed
> > > > > > >
> > > > > > >
> > > > > > > Hi,
> > > > > > >
> > > > > > > I just wonder if there are any bench figures of Hatari Falcon. I
> > > > > > > get the feeling that if you enable "cycle exact" and "prefetch
> > > > > > > mode" you get a slightly slower Falcon than the real thing. And if
> > > > > > > you disable those you get a CT2 or Nemesis speeded Falcon. Any
> > > > > > > measurements made?
> > > > > > >
> > > > > > > Just recently joined the list so my appologies if this has been
> > > > > > > discussed alot before...
> > > > > > >
> > > > > > > Br,
> > > > > > >
> > > > > > > /Thomas, New Beat