JIT FPU support problem !

[amilo3438]

Originally Posted by PeterK

Is the switch FPU: "More Compatible" disabled ? Try that, please. FPU: "More Compatible" is much slower in some cases.

Yes, by using Quickstart A4000 040 config both "More Compatible" switches are disabled, and such config is used for test.

EDIT:
I noticed something: if the demo is run with JIT-FPU disabled and changed to JIT-FPU enabled at first black screen (just after loading slider), it will play fine.

EDIT2: Just tested: Seems to work fine with WinUAE 2.6.1 and JIT-FPU enabled !!!

EDIT3: Found it: It is broken in newest WinUAE 2.7.1 b12 version !!! (In b11 it works fine.)

p.s. Just guess, maybe something in below changes broke it: (or something else)

Beta 12:

- BCHG/BCLR/BSET allowed PC-relative addressing modes, broken since the beginning. (from Hatari)

Toni will say, as usual, so maybe would be best to wait until tomorrow.

[Toni Wilen]

winuae.zip should fix it, at least eld-kakao glitches are now gone.

FPU constant value optimization introduced in 1200b3 is now disabled. Code looks correct and it works fine normally, only in some rare situations (no idea what is the trigger) it returns garbage. Probably caused by some unknown JIT side-effect that no one knows except original author..

[amilo3438]

Yeah, much better now ... thanks to all that were involved in finding solution !


All previously problematic demos now work fine except:

Machinist by Elude -> maybe some small glitches occasionally appear

We come in peace by Elude -> small problem on top lines of the screen (see below for other similar problems)

Silkcut by The Black Lotus -> still same problem (currently only solution is: JIT-FPU = OFF)


Other demos that also has problem with top lines are mentioned on Problematic AGA demos thread:
http://eab.abime.net/showpost.php?p=945375&postcount=39

[vagrant]

Trying to get Silkcut running for video capture and came across this thread.
With JIT FPU enabled there are some gfx glitches on certain effects and the demo quits during the mountain scene.
Disabling the option fixes these problems but the demo runs very slow.

Is there any chance this one could be looked at?

[Toni Wilen]

Not me. At least unless someone else finds out which instruction causes the problem.

[amilo3438]

@ED-209

Found configuration for Silkcut to run with JIT FPU enabled! (so it will not quit during the mountain scene!!!)

Quickstart Model A4000 68040 2MB Chip +64MB motherboard fast (see attached pic. and/or config. file)

DH0: demo file attached as archive (Note: to run type: silkcut.exe)

Note: tested on winuae 3220!

p.s. you can gain on speed if change in the attached config. the following values:

win32.cpu_idle=0
cpu_throttle=-440.0

and it should still work fine! (it does on my pc)


EDIT:@Toni

Found the edge case where it seems to still works fine! (at least on my pc)

attached config. "tbl-silkcutfinal.uae" but with following changes:

win32.cpu_idle=0
cpu_throttle=-422.0

With changes it seems to still work fine, but not every time! (tested 2 times)

[vagrant]

Thanks amilo. I tried applying a4000 rom / 64mb mobo & cpu throttle settings to my config and the demo crashed at mountain scene with usual 'out of memory' error.

Then I load your config directly and added my hdf to it (for some reason it didn't autodetect correct suface & sector info (!), had to edit that manually). It still crashed with the same error unfortunately.

[amilo3438]

What WinUAE version have you used?

I tested it with new WinUAE 3.2.2 -> http://eab.abime.net/showpost.php?p=1057150&postcount=1

and attached config. works fine every time on my pc!

p.s.
you can also try to create your config file:

1) Quickstart Model A4000 68040 2MB Chip (but change from 8MB to 64MB motherboard fast)
2) change CPU Speed slider to -50% (see attached picture in post #26)
3) attach the demo file "tbl-silkcutfinal.lha" with "Add directory or archive... / Select archive or plain file"
4) start the emu and type "silkcut.exe" (enter)

And enjoy the demo!

p.s. (to gain more speed)
4) save the config file from winuae
5) open the saved config file with Notepad editor and change following:

win32.cpu_idle=0
cpu_throttle=-440.0

6) start the winuae and load the saved config
7) type "silkcut.exe" (enter)

[vagrant]

Yeh, sorry forgot to mention that I'm using the latest version.

I use all those settings, and also your attached config with only change being the HDF, but it still crashes here.

I also tried increasing cpu throttle even further but it still wouldn't work.

[amilo3438]

Originally Posted by ED-209

I also tried increasing cpu throttle even further but it still wouldn't work.

Have you tried to decrease the cpu throttle: -60%, -70%, -80%, -90% (maybe your pc cpu is faster)

I guess one of these should/might work! (otherwise, I have no idea what else could help)(btw. my cpu is Pentium M @2.13GHz)

[vagrant]

Yep, that's what I meant (moving slider further to the left).
My cpu is 3.8ghz dual core.

I just tried running the demo without setpatch, as it the only difference I can see by running the archive directly instead of my hdf, still crash at the same spot.

(you can see it is your config I am running - scanlines )

[amilo3438]

Yeah, the speed of our CPU is different and generations, so that it could be the reason! (I got the same message if winuae cpu_throttle slider is set to -40% ... on -50% it does not happen on my pc)

[vagrant]

Got it! It works here on -90% throttle, any less and it crashes.
Sorry, I must have tried every option before other than 90%

The demo is dropping a lot of frames because of the throttle, but it is still smoother then with no JIT at all.. and it is nice to finally see the prod run in full

[amilo3438]

Great! (which means that the theory is now proven in practice)

So it works fine for you only on -90%! (but not on -80%)

Maybe you can gain on speed if change the values in your config file like f.ex.:

win32.cpu_idle=0
cpu_throttle=-850.0

I guess it might still work fine! (if not try with -860, -870, -880 or -890) ... (or maybe if -850 works fine you could try with -840, -830 ...)

Sorry, I have to go to sleep now! (morning is about to dawn)

[Toni Wilen]

Obviously this is a demo bug (has assumption that "shouldn't happen" because "this routine can never run that fast", bad thread syncronization, lots of possible reasons) if changing emulation speed affects behavior.

[amilo3438]

On my pc it works fine (no crashing) on cpu_throttle 0% in JIT-Indirect mode! (see on pic. changes done on config from post #26)

Is it because some difference in speed between JIT-Direct and JIT-Indirect mode, or something else!? (not sure)

EDIT:
Also, in JIT-Direct mode when "No flags" is unchecked it does not crash with cpu_throttle -40% ! (on my pc)

CPU Speed = -50% (JIT/No Flags = checked) ... no crashing
CPU Speed = -40% (JIT/No Flags = unchecked) ... no crashing


EDIT2: @ED-209
As your pc-cpu is almost 2x faster than my pc-cpu (3.8ghz vs 2.13ghz), maybe it would be nice to see if your PC not have problem when cpu_throttle=0% in JIT-Indirect mode (as shown on attached pic. below), perhaps it could prove a new theory!?

P.s.
I thought if it worked without a hitch on your PC perhaps the problem would not be in speed!?

[amilo3438]

Last Train To Danzig by Ghostown & Haujobb -> http://www.pouet.net/prod.php?which=66492

WinUAE 3.2.2

1) Quickstart Model A4000 68040 2MB Chip (change mem from 8MB to 32MB motherboard fast)
2) demo file "gtn_hjb-lasttrain-party.lha" attached via "Add directory or archive... / Select archive or plain file"
3) from CLI "lasttrain-party" (enter)

Problem: The endscroll background glitches! (see on pic. below) (not sure is it AGA or JIT problem!)

(how should it look check on above link videos)

[amilo3438]

Originally Posted by Toni Wilen

Obviously this is a demo bug (has assumption that "shouldn't happen" because "this routine can never run that fast", bad thread syncronization, lots of possible reasons) if changing emulation speed affects behavior.

Hmm, but could it not be also possible that the problem is not speed as it happens only in JIT Direct mode and not in Indrect mode!? (at least on my PC pentium M @2.13ghz)

In JIT Direct mode it doesn't quit till cpu_throttle -42%! (so this is the border area for JIT Direct mode on my PC with Silkcut demo!)

In JIT Indirect mode it works fine with cpu_throttle 0%!?

So I wonder if it would also work fine on some speedier PC in JIT Indirect mode and 0% cpu_throttling?

[Toni Wilen]

Direct is much faster than non-direct. And don't attempt to use throttling options in JIT modes to guarantee anything, JIT still can run 1000+ instructions between throttling and if timing sensitive stuff is inside that block of instructions: throttling makes no difference. It becomes only random.

[amilo3438]

Direct mode is much faster than indirect ... thanks for the explanation!

EDIT:
As I recall, Silkcut demo doesn't have problem in JIT Direct mode if JITFPU is OFF!
I wonder now if would be somehow possible to adjust only JITFPU to work in Indirect mode? (not sure how much this idea sounds impossible - read stupid... but there are no stupid questions, right?)