Request: 68030/A3000 cycle-exact

[Aegis]

Hi Toni,

as hardware gets more powerful it'd be great to be able to run more stock configurations/CPUs at cycle-exact speeds - the A500 had a few '030 accelerators for it which were useful for many applications and even some games and obviously 'fastest possible' is well, crazy fast

Would it be possible to add a cycle-exact 25Mhz '030/A3000 (and in time a cycle-exact '040 for A4000 emulation)?

[ReadOnlyCat]

Everything is possible.
The real question is "are the efforts worth the result?".

Figuring out exact timings starting from the '020 requires complete understanding and emulation of the prefetch and caching unit and they get more and more complex with each version of the CPU.
It would take a lot of painstaking work to write the corresponding code just from Motorola's documentation (if it is complete enough!) and frankly that would not be enjoyable at all.

[Toni Wilen]

As was already mentioned above, not even basic 68020 is cycle-exact yet.

Motorola documentation does not really mention anything about instruction's cycle sequencing, just some theoretical total cycle usages that are mostly useless. (I'll randomly do some logic analyzer testing, 68020 CE will improve randomly but slowly..)

Full 68020 CE is practically impossible without some undocumented information, for example multiplication and division algorithms are totally unknown and they have variable instruction cycle count.

[Aegis]

Originally Posted by ReadOnlyCat

Everything is possible.
The real question is "are the efforts worth the result?".

What that depends on what WinUAE *is* - to you, to me - and especially to Toni.

I've always seen UAE as an effort to preserve the Amiga in software (since it's gone way beyond simply running games and apps) - accurate emulation of the all the Amiga models is a given if that's the ultimate goal. But that's just me - I'm sure every user and developer has their own idea of what their 'ideal' UAE is and theirs is just as valid as mine.

Whilst it wouldn't be cycle-exact, perhaps a compromise might be to have an option to run the 030/040/060 at approximate hardware speed instead of 'fastest possible'? I know you can slow the emulation down but what % of 'fastest possible' is a 25Mhz 68030?

Thanks both of you for taking the time to reply!

[amilo3438]

Originally Posted by Aegis

I know you can slow the emulation down but what % of 'fastest possible' is a 25Mhz 68030?

Turn off JIT
Turn off CPU cycle-exact (keep only DMA/memory cycle-exact)
Go on CPU frequency, change to Custom, type 25

I think it could be close enough to a real machine.

[Aegis]

Originally Posted by amilo3438

Turn off JIT
Turn off CPU cycle-exact (keep only DMA/memory cycle-exact)
Go on CPU frequency, change to Custom, type 25

I think it could be close enough to a real machine.

Thanks for the tip! But using your suggestion, how would a 25Mhz 030 compare against a 25Mhz 040? The 68040 should be faster no?

[amilo3438]

Originally Posted by Aegis

Thanks for the tip! But using your suggestion, how would a 25Mhz 030 compare against a 25Mhz 040? The 68040 should be faster no?

Im not sure if there is any difference in speed among 030/040/060 in current emulation when run on same frequency! (On a real CPUs there is)

[Aegis]

Well that's what I was kind of getting at with my original request As I understand it (and I could well be wrong) the different CPUs available in WinUAE are basically their different instruction sets - the actual operation of the CPU is only fully emulated in the 68000 with the 68020 partially implemented.

I'm guessing the 'custom' frequency setting is in comparison to the 68000 and a 25Mhz 040 is in reality going to be considerably faster than a 25Mhz 68k. There's no way of approximating the actual speed of (say) an A4000 in UAE (well, you could find a real one and benchmark it and tailor UAE to give similar results) - for historical (and nostalgic) reasons, I think it'd be nice if there were.

[amilo3438]

Regarding a Real CPUs..

020 and 030 are ~ the same speed on same MHz

040 is ~4x faster than 030 on same MHz

060 is ~2x faster than 040 on same MHz

[Aegis]

Thanks amilo - may I ask where you got your figures from? (just curious )

[amilo3438]

Originally Posted by Aegis

Thanks amilo - may I ask where you got your figures from? (just curious )

Yeah, from experience and reading , mostly!

Cheers

[amilo3438]

Originally Posted by Aegis

Would it be possible to add a cycle-exact 25Mhz '030/A3000 ?

In mean time the attached config looks very close: (according to some of AIBB results)

AIBB v6.1 (from LSD-Tools101.ADF -> http://janeway.exotica.org.uk/release.php?id=23128)

Comparisons Against System: A3000-25 (Test Code Type: 68020, CP Math)

EmuTest ....... 0.63 (4.40 Simulated Mhz)
EllipseTest ... 1.16 (11.54 Seconds)
LineTest ...... 1.31 (1060.74 Lines/Second)
InstTets ....... 0.72 (1820338.98 Instructions/Seconds)
Writepixel ... 1.00 (4437.33 Pixels/Second)
Matrix ......... 0.99 (2.38 Seconds)
Sieve .......... 0.97 (7.36 Seconds)
IMath ......... 1.69 (3.10 Seconds)
Dhrystone ..... 0.71 (6377.55 Dhrystones/Seconds)
MemTest ..... 0.99 (7.81 MBytes/Sec Transfer Rate)
Sort ............ 1.02 (8.14 Seconds)
TGTest ........ 1.18 (1108.92 Characters/Sec)

Savage ....... 42.08 (0.08 Seconds)
Flops ........... 9.24 (4.1734 Scalar AIBB MFlops)
FMath .......... 2.60 (1.16 Seconds)
TranTest ...... 0.29 (31.04 Seconds)
FMatrix ........ 2.53 (1.78 Seconds)
FTrace ......... 5.53 (0.72 Seconds)
BeachBall ...... 2.14 (7.30 Seconds)
CplxTest ..... 1.08 (5.26 Seconds)


Btw. the configuration file has set: cpu_throttle=588.6 ... not sure does it depends on PC CPU type!? (my PC CPU = PentiumM 2.13ghz)
(If answer is yes, it means one will then need to adjust the "cpu_throttle" value in order to achieve above AIBB results!)

EDIT: List of demos that might work on A3000 -> http://ada.untergrund.net/?p=demos&y=0&c=0&h=4&v=0
(like for example "Ocs/aga Minimum Required:68020+" demos such as "Sanity-roots 2.0" and "Union-braincell" has been tested to work fine)

EDIT2: Found some more "Ocs/Ecs Minimum 68020+" demos:
Grind by Dekadence & Accession -> http://www.pouet.net/prod.php?which=55576 (tested to work fine)
Coredump by MoraNet -> http://janeway.exotica.org.uk/release.php?id=27489 (works but has a problem on one part!?)

[Aegis]

Hey! Thanks for doing something I was too lazy to do myself

I suspect the CPU throttle is relative to whatever host processor you're using (i7 here) - I'll test out your config to confirm (and that was kinda my thinking behind having this feature in WinUAE since UAE would likely be a better judge of 'correct' CPU speed for 030/040/060 relative to the host than I am )

[amilo3438]

Originally Posted by Aegis

I suspect the CPU throttle is relative to whatever host processor you're using (i7 here) - I'll test out your config to confirm

Yeah, that would be nice to see is it relative to host CPU or not! (thx)

Btw.
In addition, found a list of tested demos on a real A3000-25 with recommended boot method...
http://www.amigareport.com/ar139/p1-8.html

[bebbo]

Best option for cycle exact performance testing is the real Amiga with a real cpu.

But what's the best option to measure almost real performance without using real hardware
today for 68000, 68020, 68030 and maybe 68040?

WinUAE? Vamos? Any ideas?