Re: [hatari-devel] Hatari on constrained devices

[Nicolas Pomarède <npomarede@xxxxxxxxxxxx>]

Le 16/01/2022 à 16:18, Christian Zietz a écrit :
> Thomas Huth schrieb:
>
> > It's not measurable on modern CPUs.
>
> I would agree to that.
>
> Linux developers seem to look down on other platforms, because Linux has
> the incredible "perf" tool. So, why not use it, I thought?
>
> This is on a Raspberry Pi 400, certainly a less powerful CPU than in a
> current PC. Somehow inspired by the command used during the EmuTOS
> tests, I ran...
>
> SDL_VIDEODRIVER=dummy SDL_AUDIODRIVER=dummy perf stat -e cycles -r 10
> hatari --log-level fatal --sound off --fast-forward on --run-vbls 400
> --fast-boot on --tos etos512de.img
>
> ... to measure the amount of ARM CPU cycles spent, without any video or
> audio output overhead. These are the average values and standard
> deviations for 10 repetitions ("-r 10") respectively:
>
> ENABLE_SMALL_MEM=0
>   3,333,167,644      cycles   ( +-  0.45% )
>
> ENABLE_SMALL_MEM=1
>   3,267,861,361      cycles   ( +-  0.53% )
>
> Therefore, ENABLE_SMALL_MEM=1 even seems to be (very marginally) faster
> in this test!

Hi

just on the method you used to measure speed difference , note that 
video output is not disabled, so you might get some video driver 
interfering with your test, adding some fixed delay, pause, sync or others.
I would suggest to add "--disable-video 1" to really remove SDL's audio 
*and* video output, which will give a much better idea of what the 
emulation is really using , with as little SDL/OS interference as 
possible (also you can remove --fast-forward in that case and use the 
--benchmark option)

eg :

hatari --benchmark --sound off --disable-video 1 --run-vbls 8000 
--fast-boot on --tos etos512de.img

Note that 400 VBL is a rather short time, ~8 sec of emulated machine, 
you might push it a little higher.

as for the result itself, I can't say if small mem makes any big 
differences.

Nicolas