Imagine (JIT) FPU problem

[bernd roesch]

about the fpu to test. I remember there is imagine raytracer for amiga and this never work with JIt and FPU. maybe with new FPU code it work. but seem i delete my imagine install from my amiga hardfile, so i can not test. maybe i have it on older backup

[PeterK]

Originally Posted by bernd roesch

imagine raytracer for amiga and this never work with JIt and FPU.

Hello Bernd,
what means never? When did you try that for the first time, and why no bug report?

Looking onto my harddisk I found some Imagine packages v5 and an update v5.13, but can't remember that I've ever installed or tested Imagine. Do you want to have them?

[bernd roesch]

Originally Posted by PeterK

Hello Bernd,
what means never? When did you try that for the first time, and why no bug report?

Looking onto my harddisk I found some Imagine packages v5 and an update v5.13, but can't remember that I've ever installed or tested Imagine. Do you want to have them?

I have bug report, this was long ago year 2000 or so, when amithlon test and bernie do the jit. but it was hard to fix, imagine do some change on float precision and rounding tricks. in the jit was a define to set it to 80 bits, but this do not work. I have test now newest winuae.7z. problems can see in chaospro (a 3d fractal generator) when choose the built in preset deep space and move the 2D view. the 3d view get render errors. when jit is disable work better.

the imagine is too on the eab ftp server. but if i find my old backup it is more easy to test, because i have all installed and some projects done

[Toni Wilen]

Originally Posted by bernd roesch

problems can see in chaospro (a 3d fractal generator) when choose the built in preset deep space and move the 2D view. the 3d view get render errors. when jit is disable work better.

But does 80-bit host mode (without JIT) work better than old 64-bit mode? Is there differences between 80-bit mode and softfloat mode?

(too lazy to test just yet..)

[PeterK]

I'm not sure if Bernd is talking about some sort of digital dust or snow when he says "3D render errors", but that seems to disappear when Jit is switched to indirect mode. Jit off and softfloat makes not much fun, it's terrible slow.

Update1: It's not completely perfect in Jit indirect mode but much better or at least different. This strange line from the lower right area to the upper left side looks like a rounding bug. Maybe some intermediate results are getting wrong?

Update2: In old WinUAE 2.4.0 these strange lines are much worse, and they still exist even in softfloat mode with beta9.

[bernd roesch]

the strange lines i see in 64 bit host, 80 bit host and softfloat 80 bit when disable JIT. when enable JIT there are much mor display errors. press and hold mouse button to move the 2d view much. you see that it happen soon that only a yellow screen is display

[PeterK]

My assumption is that ChaosPro is buggy. WinUAE may just trigger these bugs. Probably some buffer or memory access problems. There is no area that can't be calculated and displayed correctly, the issues only occur when you move the 2D view around. You can close the 2D window and then restart the calculation and everything appears perfectly.

[Sinphaltimus]

I loaded this up on a real Amiga equipped with a Vampire 500v2+ and did not immediately see any issues. Movin the 2D screen around after several redraws and i think I began seeing similar issues. Granted the Vampire FPU is not 80-bit. I think it's in the area of 64. Also, I had to use my phone to take the pics...

Here is the first render with no 2D movement...


Slight movement and a little graphical "not so smoothness" appearing upper right quadrant.




Some more movement and there we start to see some pretty bad anomolies appearing upper right quadrant...



And here we see the worst one yet left side where an actual out of place line appears like in other screenshots posted prior to mine...




If you have imagine 5.0 files to share that i can test I'm willing to do that as well.

[PeterK]

Thanks Sinphaltimus,

the supported precision (64 or 80 bit) should not be responsible for these issues. Even with 64 bit you can calculate already most parts, but with 80 bit it might be possible to zoom a little deeper, which I didn't try yet. It's good to know that not only WinUAE shows these strange effects.

Back to the topic: Imagine!
Unfortunately, I have no ready to use objects for Imagine 5.x for tests. Imagine seems not to have an easy and intuitive concept and I don't like programs like that. Bernd said that there is a CD image on the EAB server, which I didn't try yet.

[hexaae]

Originally Posted by PeterK

My assumption is that ChaosPro is buggy. WinUAE may just trigger these bugs. Probably some buffer or memory access problems. There is no area that can't be calculated and displayed correctly, the issues only occur when you move the 2D view around. You can close the 2D window and then restart the calculation and everything appears perfectly.

Yes, CP is buggy: I use it also in my WinUAE env and the only way to make it work without rendering errors (and big windows > 800x600) is disabling the software buffers (= select "No Buffer"):




"16Bit-Int buffer" and "IEEESP buffer" cause artifacts and issues dragging the fractals around...

[bernd roesch]

I find imagine. it is version 4 but objects and films are in. problem is, it open on a interlace screen, but the interlace screen is so small and i find not the size to fullscreen switch on new winuae and that it show the interlace screen that not flicker and is 2* larger. what settings i need to get it larger ?

[PeterK]

Try Filter => AutoIntegerScaling

[bernd roesch]

I test imagine 4/5 with new winuae and JIT. it work ok now


only i notice with sysspeed. when use winuae older or same 3.5.0 then FPU mips show around 2000. with 3.6.0 FPU mips show only 35. more precise i can not tell, because i have not all winuae store. or can i donwload somewhere oher versions and beta ?. settings are the same. but in hd-rec reverb benchmark or other newer winuae version FPU is not slower.

slowdown seem only happen in sysspeed mips test

[Toni Wilen]

Sysspeed FPU tests looks very stupid. It runs simple loop thousands of time which does fadd, fsub, ftst and fneg. With all FPU registers containing NaN value!

Modern FPU hardware probably uses microcode (=slow) for special cases that don't need to be fast and are rare.

[PeterK]

Some benchmark results of SysSpeed, SysInfo and AIBB are just crap or out of range. Don't expect any meaningfull results with programs from the early 90's on fast PC's.

There are a few older WinUAE versions back to 3.0.0 on the WinUAE page, and I also have a collection of many old versions since 0.822 on my old HDD.

http://www.winuae.net/download-old-versions/

[PeterK]

Ok, I'm hijacking this thread now with Elena Novaretti's ZoneXplorer...
There is a difference between SoftFloat or Jit FPU and non-Jit in the ZX module DistMaps/SpaceConcept. With SoftFloat and Jit FPU the outer background appears white, but in non-Jit mode all the background is black.

Can somebody with a real FPU check out on an Amiga what's right and what's wrong?

It's hard to say which FPU instructions are responsible for this difference. There are hundreds of floating point functions in ZoneXplorer and the modules. This bug already exists in older versions of WinUAE, like 3.6.1.

Update: In 2.4.0 the Jit still has a completely black background and in 3.0.0 it seems already to be broken. If this bug was introduced between these 2 versions then SoftFloat is also affected by this issue.

[PeterK]

Another one bites the dust: FlashMandelNG 1.1 doesn't like PAL(crashes), SuperPlus and HighGFX, but works on UAE screenmodes. (set the screen size to the default like 800x600 or 1024x768, and colors to 256)

http://aminet.net/package/gfx/fract/FlashMandelNG

I'm not sure if this is caused by the FPU, could also be something else. Also happens with all older versions of WinUAE.

[Toni Wilen]

Sorry but simply guessing if some application works correctly or not is not enough and then assuming it may have something to do with emulation. I don't debug applications until it is 100% confirmed (because it is almost always application bug EDIT: and there are far too many Amiga side differences that can cause it. Just installing the software in another WB may or may cause same problem!)

Crash with JIT enabled = means nothing. Crash dump with JIT enabled = useless dump file.

[PeterK]

Of course, this can be caused by FlashMandel (or CrashMandel), but I have no real Amiga anymore to compare the behaviour.

With SoftFloat on PAL it just stops working, no crash.

[Toni Wilen]

Check log (JIT crash log showed lots of invalid memory access = it most likely is application bug or some related lib is incompatible), enable MMU emulation and use enforcer (or other similar debugging tools). Make sure stack is not too small. The usual Amiga debugging.