Rendering the SAME vrscene on WinXP64 and Fedora15 producing different output

[ultrasonic]

Hi Guys,

are you aware that image output on linux and windows is slightly different? If yes, can u explain why? I thought I can mix OSes on my renderfarm...
I used the SAME vrscene on both oses...rendering with standalone build 16633 (both for maya2012)

Do you think Fedora vs. MacOS will also be different?

Here are two examples....(ignore the overexposure)

If you like here are the exr's:
http://fullblownimages.com/download/diff_lnx_win.zip

cheerio
Oliver

[Dariusz Makowski (Dadal)]

Its not different.

The GI is at very low settings causing samples to distribute differently causing ur car reflections to be different... Pop up ur settings and all be fine...

Take care.

[ultrasonic]

I am rendering the VERY SAME vrscene, so the images should look identical...this was more a question to the developers of vray...

[Dariusz Makowski (Dadal)]

I know u render that same scene. But each time u render scene vray is casting random number of samples. If u set GI to very low the samples, the map will be very noisy and each time u render that same scene the effect would be slightly different.
Look at the tree in left upper corner - the GI noise is changing from image to image...

Render the scene 3x on windows and u'll see that each image will be slightly different...

If u want to use such low settings use 1 pc to prebake all GI and then use it from file on all nodes...

Take care, Let me know how the test went... I'm curious if its really GI or something else...

[ultrasonic]

Hey Dadal,

you are 50% right in my case....i just tested some more setups....win_maya2009, win_maya2012,fedora_2012 and another fedora15 on vritualBox.
Here is the updated archive of the exrs for comparison:
http://fullblownimages.com/download/diff_lnx_win.rar

The files named win2009 and win2012 really show only some minor difference in the GI.
As do the linux exrs. they also have GI differences only. If you want i can make a test without GI on top... EDIT: attached it below
But if you compare lnx vs windows renders you will notice some GEOMETRIC problems e.g. in the lower left of the image. The motion blurred grass is definitely different on both OSes.

Are you with me in this case guys?
EDIT: Here are 4 more images rendered without GI....2 images rendered from the same vrscene on the same windows machine and again on one linux machine.
www.fullblownimages.com/download/diff_lnx_win_noGI.rar

[Dariusz Makowski (Dadal)]

Heya

Interesting results !

right No GI Images - I only see difference in Noise which is quite interesting... maybe pop up sampling and ull get equal results... ?

the GI samples.
Hard to say on such low settings - would it be possible if u could up the GI settings maybe ?
From other side I've notices in reflection of car - door section there is tree - its Different between fedora and win - like fedora dont have direct lighting on it...
This leads me to 2 conclusions.

1. Fedora cant access texture to lit the object
2. There might be difference in machines in terms of ram and so on - based on that the scene geometry and visibility of object, maybe could change hide/change whats in background causing the tiny difference ? - Not sure if thats possible I only use max and maya from what I remember is rly... nuts.

[vlado]

It looks to me more that mostly the noise pattern is different, which should not be that difficult to unify across the OSes. The reason is that the compilers that we use for Linux and Windows are different (gcc vs Intel C++). If you are using VRayFur for the grass, which also relies on random numbers, said compiler differences might be enough to cause slight differences in the fur strand distribution as well.

In general though, it is not a good idea to mix OSes for rendering; because of the different compilers, they are simpy not running quite the same code.

Best regards,
Vlado

[ultrasonic]

Hi Vlado,

the compiler might be an issue...but just FYI i am using no Vray Fur for the grass....it's an old scene where i converted maya pfx to proxies...so the only random number that may cause problems maybe the dmc noisepattern....

Would be really nice to have same random numbers on WIN, linux and macos...
EDIT: I can now confirm that linux and MacOS look identical...so the compile is pretty sure the reason.

[t.petrov]

Hello,

I've just committed a fix for this problem. The fix will be in the next nightly. Can you try
it and then report if the image looks the same on both Windows and Linux?

Best regards,
Teodor

[ultrasonic]

I tested again it looks very promising...need to further evaluate this in upcoming projects...
There are still some slight differences in the noise pattern..is that intended behaviour when one renders the same vrscene on two different machines? If yes, then what is the reason for this?

On the other hand i compared rendertimes between linux and windows and it seems that windows is a lot faster...the two machines i tested on were both 2.6ghz quadcores and rendertimes varied by more than 10 percent...2300 vs 2700 seconds...couldn't test it on the exact same machine but it is really worrying me...any ideas if this could also be caused by the compiler?

regards
Oli

[t.petrov]

Unfortunately, the fix has been reverted, because it broke the noise pattern very badly, we will try to fix it correctly.
The revert happened in revision 16955. Revisions between 16876 and 16954 are broken.

As far as I know Intel's compiler is 10-20% faster than MS' compiler, GCC and VisualC++ are almost identical in performance.
We've not tried Intel's compiler on Linux, because we are afraid of Symbol collisions.

[ultrasonic]

Unfortunately, the fix has been reverted, because it broke the noise pattern very badly, we will try to fix it correctly.
The revert happened in revision 16955. Revisions between 16876 and 16954 are broken.

Ok....but i tested it in build 17038 and it looked ok to me...so you say that the fix was already removed there? Strange....

As far as I know Intel's compiler is 10-20% faster than MS' compiler, GCC and VisualC++ are almost identical in performance.
We've not tried Intel's compiler on Linux, because we are afraid of Symbol collisions.

Hmmm...10%-20% is about the decrease i noticed...since speed is everything i would beg u to look how you can improve mac & linux builds...i have no idea how hard this is, but it is kind of a NOGO for me to switch to linux when i know that all my renders are going to take that much longer...would it be possible to have two different builds for linux using the intel compiler AND gcc as a fallback? I think houdini is offering a similar option.

cheers
Oliver

[alainfx]

As far as I know Intel's compiler is 10-20% faster than MS' compiler, GCC and VisualC++ are almost identical in performance.
We've not tried Intel's compiler on Linux, because we are afraid of Symbol collisions.

On which side is VRay on Mac ?
Windows speed or Linux speed ?
Thanks.

[t.petrov]

GCC speed + slower processors most of the time

[alainfx]

Thank you.
Windows 7 is so good now, no need to look elsewhere anyway.