JIT FPU support problem !

[vagrant]

Amilo, I tried running Silkcut with JIT in "indirect" mode.
This time the demo needed =>70% CPU throttle to run without crash, however it was running quite slowly and lots of gfx glitches happening.

Also about "Last train to Danzig":
I captured this demo when it was first released without any real problem? It was during 3.2 beta phase. Config was just standard 020+JIT+Z3 fast.
My capture here: [ Show youtube player ]

[amilo3438]

Some more testing with Silkcut demo: (on my PC 2.13ghz)

1) not crashes (cpu_throttle 0%) (SySInfo_v4.0: Mips=228.02, MFlops=236.98)

comp_trustbyte=indirect
comp_trustword=indirect
comp_trustlong=indirect
comp_trustnaddr=indirect

2) not crashes (cpu_throttle 0%)

comp_trustbyte=direct
comp_trustword=indirect
comp_trustlong=indirect
comp_trustnaddr=indirect

3) not crashes (cpu_throttle 0%)

comp_trustbyte=indirect
comp_trustword=direct
comp_trustlong=indirect
comp_trustnaddr=indirect

4) crashes (cpu_throttle 0%)
not crashes if cpu_throttle=-350.0 (-35%) (SySInfo_v4.0: Mips=512.61, MFlops=183.56)

comp_trustbyte=indirect
comp_trustword=indirect
comp_trustlong=direct
comp_trustnaddr=indirect

5) not crashes (cpu_throttle 0%)

comp_trustbyte=indirect
comp_trustword=indirect
comp_trustlong=indirect
comp_trustnaddr=direct

6) crashes (cpu_throttle 0%)
not crashes if cpu_throttle=-450.0 (-45%) (SySInfo_v4.0: Mips=406.70, MFlops=146.40)

comp_trustbyte=direct
comp_trustword=direct
comp_trustlong=direct
comp_trustnaddr=direct

7) crashes (cpu_throttle 0%)
not crashes if cpu_throttle=-50.0 (-5%) (SySInfo_v4.0: Mips=225.57, MFlops=237.53)

comp_trustbyte=direct
comp_trustword=direct
comp_trustlong=indirect
comp_trustnaddr=direct


Interesting cases for Silkcut demo when not crashes: 1), 4), 6) and 7)

In cases 1) and 7) SySInfo_v4.0 shows almost same speed, but in cases 4) and 6) it shows much higher in Mips with lower in MFlops !?

Could that mean that the case 4) is the best choice (regarding to much higher in Mips) or maybe cases 1) or 7) are better choice (regarding to higher in MFlops)!?

EDIT:
Comparing again cases 4) and 7) by running the Silkcut demo, obviously the case 4) run much smoother on my PC!!! (so case 4) is the winner)

EDIT2:
Let test & compare again cases 4) and 7) with sysinfo_v4.0 Dhrystones:

case 4 , cpu_throttle=-300.0 , sysinfo_v4.0 Dhrystones=557106 , Silkcut demo crashes!!!
case 4 , cpu_throttle=-350.0 , sysinfo_v4.0 Dhrystones=491279 , Silkcut demo still works!!!

case 7 , cpu_throttle= 0.0 , sysinfo_v4.0 Dhrystones=226448 , Silkcut demo crashes!!!
case 7 , cpu_throttle=-50.0 , sysinfo_v4.0 Dhrystones=217817 , Silkcut demo still works!!!

So, can we finally dismiss the impact of speed to the Silkcut demo crashing !? (491279 is much greater than 226448)

[amilo3438]

Originally Posted by ED-209

Amilo, I tried running Silkcut with JIT in "indirect" mode.
This time the demo needed =>70% CPU throttle to run without crash, however it was running quite slowly and lots of gfx glitches happening.

This is a little weird as your PC cpu is only 2x faster than my! (on my PC in Indirect mode it runs fine with 0% CPU throttle)

Maybe you can try with some changes like f.ex. my new config case 4), but for you it will probably work with cpu_throttle=-700.0 i guess!
(OK, I have attached the my new config "case 4" for you to test - check below )(Note: It has already changed the cpu_throttle=-700.0 !!!)

Originally Posted by ED-209

Also about "Last train to Danzig":
I captured this demo when it was first released without any real problem? It was during 3.2 beta phase. Config was just standard 020+JIT+Z3 fast.
My capture here: [ Show youtube player ]

And I tested it with Quickstart 040+JIT+fast 64MB config! (thats is interesting)

Can you attach your working config for "Last train to Danzig" demo to test!? (I still have the same problem - missing "fog" effect at beginning and endscroll)

[amilo3438]

Some more testing...

Analysis of the JIT cashe size on the emu speed in Dhrystone!

Tested config: WinUAE 3.2.2, Quickstart A4000, 68040, 3.1, 2MB Chip + 8MB Fast (CPU Idle 0) (on PC=Pentium M 2.13ghz)


JIT-Direct

cachesize=8192 , Direct , sysinfo_v4.0 Dhrystones=663743

cachesize=20 , Direct , sysinfo_v4.0 Dhrystones=620558

cachesize=19 , Direct , sysinfo_v4.0 Dhrystones=10632

cachesize=12 , Direct , sysinfo_v4.0 Dhrystones=5913

cachesize=<10 -> JIT:Serious error:Could not allocate block pool!


JIT-Indirect

cachesize=8192 , Indirect , sysinfo_v4.0 Dhrystones=226201

cachesize=22 , Indirect , sysinfo_v4.0 Dhrystones=200277

cachesize=21 , Indirect , sysinfo_v4.0 Dhrystones=8632

cachesize=20 , Indirect , sysinfo_v4.0 Dhrystones=9033

cachesize=19 , Indirect , sysinfo_v4.0 Dhrystones=9466

cachesize=12 , Indirect , sysinfo_v4.0 Dhrystones=5562

cachesize=<10 -> JIT:Serious error:Could not allocate block pool!


p.s. Also, the post #42 has been updated with EDIT2 !!!

[vagrant]

I tried the latest config file for Silkcut.
This time it run complete without any "glitches", although it is dropping lots of frames in the more cpu-intensive scenes.
Any lower than 70% throttle and the demo quits at mountain scene, which is a shame as it seems a good amount more cpu is required for a full framerate capture.

Last Train To Danzig:
Yep this works fine on my default RTG/JIT config. I will attach, but be warned it is far from 'quickstart' settings

[amilo3438]

Originally Posted by ED-209

Last Train To Danzig:
Yep this works fine on my default RTG/JIT config. I will attach, but be warned it is far from 'quickstart' settings

Thanks, so it works fine with RTG/JIT and missing the "fog effect" on AGA/JIT, right !? (Btw. I have tested it only on AGA)

Hmm, I didn't find anywhere that is said "for RTG only", so something might be wrong with WinUAE AGA emulation than, I guess... of course, this needs to be checked/tested on a real AGA 68040/60 Amiga too!

Originally Posted by ED-209

I tried the latest config file for Silkcut.
This time it run complete without any "glitches", although it is dropping lots of frames in the more cpu-intensive scenes.

One thing I forgot to mention about this "experimental" configuration files...
It needs to be changed only with some editor like f.e. Notepad, otherwise if is changed via WinUAE menu it will reset to default settings! (so its good to check that the mentioned values in the config file are not changed before loading/starting it on emu!)

[vagrant]

I double-checked Silkcut to be sure.. This time I load your config, then manually edit throttle with "uae-configuration cpu_throttle -500" prior to running the demo, and it still exit at mountain scene.

Last train to Danzig:
No the demo isn't RTG, this is just the main config I use
I played around with the a4000 quickstart config and it looks like the problem is cause by using "motherboard fast"! Try switching to Z3 instead and it should work

[exolon]

Hi ED!
About Silkcut - of course, this can be done!
Check this http://thecompany.pl/gra/The+Black+Lotus+Demography
v2.6 [2014-02-23]
+ Silkcut is 100% working (without disabling JIT!).

[Toni Wilen]

[above post]

You don't talk about that site in this (sub)forum, also it is not using official WinUAE version. This topic ends here, future replies will disappear.

---

[previous posts]

This is quite pointless. You are trying to measure something that does not exist. Benchmarking JIT is meaningless, trying to adjust speed is mostly pointless.

[amilo3438]

Originally Posted by ED-209

Last train to Danzig:
No the demo isn't RTG, this is just the main config I use
I played around with the a4000 quickstart config and it looks like the problem is cause by using "motherboard fast"! Try switching to Z3 instead and it should work

Thanks, I forgot to try the demo with Z3 mem... so below is my new test findings:

Used config: Quickstart A4000, 040 , 3.1 , 2MB Chip + 8MB Fast + Z3 Fast: 32MB

Z3 mapping mode:

Automatic (*) ... The "fog" effect is missing!
UAE (0x10000000) ... The "fog" effect is missing!
Real(0x40000000) ... It works fine!

Originally Posted by ED-209

I double-checked Silkcut to be sure.. This time I load your config, then manually edit throttle with "uae-configuration cpu_throttle -500" prior to running the demo, and it still exit at mountain scene.

Regarding the Silkcut demo, let a little math:

If on my PC 2.13ghz the case 4) works fine with cpu_throttle=-350.0

3.8ghz / 2.13ghz = 1.784 -> -350.0 * 1.784 = -624.4

so maybe on your PC 3.8ghz it might still work fine with f.e. cpu_throttle=-650.0! (I guess)

(obviously the cpu_throttle=-500 is then "too fast" for 3.8ghz, right?)

[vagrant]

exolon: I will talk to you soon via email

Originally Posted by Toni Wilen

Benchmarking JIT is meaningless, trying to adjust speed is mostly pointless.

I was kind of waiting for this, since toni say before that throttling JIT is undependable and cause random result

amilo:
confirm Last Train to Danzig;
automatic / real are working here.
UAE = no fog.

[amilo3438]

Originally Posted by ED-209

I was kind of waiting for this, since toni say before that throttling JIT is undependable and cause random result

Well, at least in case of Silkcut demo it shows very usefull as it helps to run the demo with acceptable speed and without quitting on mountain part!

Btw. regarding the benchmark results, yes, quite pointless as absolute values, and its only there to show feeling of speed for each test case!

[amilo3438]

JFI...

Klone by Dual Crew & Shining -> http://www.pouet.net/prod.php?which=1562

An another case, similar to Silkcut demo, that it is possible to make working on JIT with decreasing "CPU Speed":

JIT-Direct (CPU Speed = -80% - 0%) - after 1st part got black screen and music play!
JIT-Direct (CPU Speed = -90%) - it works fine!

JIT-Indirect (CPU Speed = -40% - 0%) - after 1st part got black screen and music play!
JIT-Indirect (CPU Speed = -50%) - it works fine!

Note: The above values are valid only for my PC! (my PC = Pentium M 2.13ghz)

p.s. Tested on WinUAE 3220, QS A4000 040!


EDIT: More examples...

Massive by skarla demo ...

JIT-Direct (CPU Speed = -60%) - it does not flash colour/glitches!
JIT-Indirect (CPU Speed = -50%) - it does not flash colour/glitches!

Otherwise, it might occasionally flash colour/glitches in any of Massive demo parts!

Machinist by Elude demo ...

Above values also help to reduce/eliminate glitches on f.e. "vulcano cave" part in Machinist demo!!!


Perhaps one could ask the question:

If indeed all of the above mentioned demos are sensitive to CPU speed ? (Or could still be some unknown problem with JIT emulation ?)

Of course, if the problem is with JIT, it probably will not be easy to solve.

So there is a temporary solution -> if a similar problem occurs, just play with WinUAE "CPU Speed" slider, may help!

[amilo3438]

JFI... found an interesting case!

Shameless by Spaceballs -> http://www.pouet.net/prod.php?which=12667

WinUAE3220 (config. attached) ... type: shameless.exe (from CLI)

If attached config. where JIT=Direct, after a while it will first crash into debbuger window (attached)! (the 1st .dmp file) (but if CPU Speed=-70% it will run fine)

If same config. but JIT=Indirect it also crashes but will open a crash window! (the 2nd .dmp file) (but if CPU Speed=-50% it will run fine)

If same config. but JIT=OFF it will run fine!

Note: Above CPU Speed % values are valid only for my PC! (Pentium M 2.13ghz)


EDIT: In addition.. same config. tested in older WinUAE versions:

2.0.1 (2009.12.23) = stops but does not crash
2.3.2 (2011.06.02) = got a guru! (Error: 8000 000E Task: 0781DF80)
2.3.3 (2011.09.18) = stops but does not crash
2.4.1 (2012.05.09) = stops but does not crash
2.5.1 (2012.12.22) = got a guru! (Error: 8000 000E Task: 0781DC40)
2.6.0 (2013.05.16) = crash into debbuger window

[amilo3438]

From WinUAE 3.2.0 we got 64-bit compatible 680x0 JIT engine! (which may be different than the 32-bit, I guess)

So, it would be good to test/check some of JIT 32-bit possible problems also on WinUAE JIT 64-bit engine! (unfortunately my PC CPU is only 32-bit)


Recommended test configuration would be:

WinUAE 3.2.2 64-bit, QS A4000 040, 2mb Chip 8mb Fast +64mb Z3


Needed demos to test/check would be:

Silkcut by The Black Lotus (see post #1 for link)

Klone by Dual Crew & Shining (see post #53 for link)

Shameless by Spaceballs (see post #54 for link ... Note: this one could crash badly on 32-bit JIT)


@ ED-209
What WinUAE version have you been using for Silkcut demo test; 32-bit or 64-bit? (I guess your PC CPU is 64-bit capable. right?)

[Toni Wilen]

I repeat: don't try to find and explain differences that are only caused by different speeds.

Shameless corrupts the stack, I am not sure if JIT can recover from that. Lets see what Frode says.

[amilo3438]

Originally Posted by Toni Wilen

I repeat: don't try to find and explain differences that are only caused by different speeds.

OK, just an testing, do not worry! (hope that it will now not afraid ed-209 to answer)

Originally Posted by Toni Wilen

Shameless corrupts the stack, I am not sure if JIT can recover from that. Lets see what Frode says.

Thats interesting, as it does not crash if "CPU Speed=-50%" on my PC! (Ok, lets see what Frode says!)

[Toni Wilen]

Originally Posted by amilo3438

Thats interesting, as it does not crash if "CPU Speed=-50%" on my PC! (Ok, lets see what Frode says!)

It only means demo gets too badly confused if CPU is too fast. I can duplicate demo crash (due to stack corruption) without JIT.

[amilo3438]

Originally Posted by Toni Wilen

It only means demo gets too badly confused if CPU is too fast. I can duplicate demo crash (due to stack corruption) without JIT.

Ok, so its possible to crash the demo also without JIT! (thx for the answer)

If I get it right, it means that the problem was in the demo and not emu!

p.s.
Really bugging me all the time how it is possible that an excessive speed may have an effect on the execution of the program !? (so the dilemma is finally resolved)

p.s.2
So, if the "demo gets too badly confused if CPU is too fast" could mean that the one who programmed it was also confused (ie.was in a hurry), right!?

No, according to number of AGA demos found, it cant be said that the coder was unexperienced!
Have been checked also some other demos from the same coder and did not find any problem! (all worked fine on JIT)

So, the Shameless demo crash was possible to duplicate without JIT on emulator - wonder if would be possible to duplicate on a real machine too!? (if one with a such speed would exist)

[vagrant]

Originally Posted by amilo3438

What WinUAE version have you been using for Silkcut demo test; 32-bit or 64-bit? (I guess your PC CPU is 64-bit capable. right?)

Actually I had been using 64-bit version regularly since JIT support was introduced and I think Last Train and others where recorded using it.
But I have since switched back to 32-bit, because aside from everything seemingly running exactly the same, I am paranoid that the 64bit version is not as compatible as 32bit

Toni: Do you regard 32bit version still the most compatible, all modes considered? If there should not be real difference between the two, then I will switch back to using 64bit indefinitely.

As for Silkcut, Klone & Shameless - they all perform identical in both builds. Same speeds, same errors.