GPU Rendering difference between Cuda and OpenCL? One is softer?

[Christopher-lee dos Santo]

So I've searched and I have a headache from reading threads that have not been able to answer my question.

I have two machines here that I have been rendering Prores 422 HQ proxies from for a feature edit. The film was shot on Dragon 5k.

Computer A has Cuda support and did 99% of the transcoding to Proxies. Computer B does not have Cuda support, but does have Open CL. This machine is only for the edit.

During the edit, we decided to change a basic look on a few shots, so we made the changes in redcine x using Computer B. After exporting the new looks into the same Proxy 1080p Prores 422HQ and comparing the two shots, it looks like the OpenCl exports from Computer B are softer?

Currently doing a CPU only export from Computer B to see if it makes any difference, but it is slightly worrying.

Is this possible or am I just blind?

 

[Phil Holland]

I haven't experienced visual differences when testing CUDA versus openCL, I've found better performance out of OpenCL in general though. I would double check the look and export settings, especially the debayer quality in the export preset. Also, on the system that supports both crank out encodes that use both forms of acceleration. Outside of that I can't think of anything else that would effect this.

 

[Christopher-lee dos Santo]

Thanks Phil. I too am confused. The Debayer settings are identical. Both are set to 4k - 1/2 Premium. That was the first thing I checked. No unsharp mask or any sharpening was applied to the footage on either machines. The proxy files are 1080p.

Here are some screengrabs.

First 1 is Computer B with OpenCL - This pumps out a prores 422 proxy in 2 minutes



This second image is from Computer A, which takes 35 minutes for a 422 Proxy.



Full res - Click on the link below, then click on the magnifying glass to compare full screen images.
PC B - Soft - http://s853.photobucket.com/user/Cr...enShot2014-11-22at182010_zps3ed562aa.png.html
PC A - Sharp - http://s853.photobucket.com/user/Cr...enShot2014-11-22at182734_zpsafcb0c5b.png.html

Computer A has a much sharper image. Ignore the colour differences, one uses RedGamma4 and the other is RedLogFilm. Export settings are identical.

I am blown away here at the difference. I assure you, this is not human error. I have checked over the settings between the two machines and they are both the same. Even exporting out with the same gamma profile gives the same results. Computer B is softer than Computer A.

There has to be something going on during the decode. Could it be graphics driver incompatibility?

 

[Mike P.]

What output scaling software are you using (file>export>options tab)? Bilinear is faster but smoother/less detail... Sinc is sharpest (I think)...

Also, the grade/colour correction/contrast differences will have apparent sharpness differences, so unless they're the same, you're really not comparing apples to apples. Maybe they look different a full res, but other than contrast, the two images you posted look to be pretty much the same sharpness to me...

 

[Christopher-lee dos Santo]

You really have to look at the pictures I uploaded in fullscreen. Also, try looking at them one after the other, maybe two browser tabs. Unfortunately, looking here on Reduser will show no difference.

I have no idea what is going on, but when ingesting the files into our NLE, we see the difference straight away. I use Lanczos3 as the software scaling. Should I rather try something else? Thanks for the help guys.

Will continue testing and let you guys know what I find out.

 

[Mike P.]

Lanczos3 is fine; I just wanted to make sure it was the same on both renders.

I was looking at them fullscreen... I still think you should apply the same look, as contrast (micro and otherwise) plays a huge part in perceived sharpness.

 

[Christopher-lee dos Santo]

Yeah, I have applied the same look. Sorry about the uploaded pics being of different gamma settings. I guess that just confused the issue at hand. Looking at the same looks, there is still a massive difference. Very weird.

 

[ericyoung]

I think you need to upload two pictures with the same look, or it's really hard to work out what's a contrast difference and what isn't. I agree there does appear to be a difference in noise/texture rendering.

There's also an x-y offset in the two pictures which doesn't help the comparison.

 

[Christopher-lee dos Santo]

Ok... This is strange... but after pulling my hair out. If I render out from Computer B using a Debayer setting of 1/2 Good as opposed to 1/2 Premium, I get an identical data rate (188.39Mbit/s) and the image is sharp again. Now the shots between computer A and B are identical. What gives?

Is Debayer 1/2 good better?


Here is 1/2 Premium debayer




Here is 1/2 Good debayer




1/2 Good debayer is sharper? What gives?

 

[Mike P.]

It's perceived sharpness; there is actually more moire/aliasing in the 1/2good render, but when scaled down/back to 1080, it actually looks sharper even though there are less pixels. Again, micro-contrast... Look a the writing on the dry-erase board; it looks "clearer" because there are less intermediate (transition) pixels between the black marker lines are at their thickest, and the whiteboard's white is at its cleanest.

You've seen it before; it's why a camera like the Alexa "looks" just as sharp as 4k/5k, but is only 3.2k. At 1:1, or on a 4k display, or blown up to 70mm projection, you'll see it's shortcomings... On a 1080p monitor, not so much. It's also why a lot of pundits say "4k/resolution doesn't matter"... It's half-true; yes it looks just as good downressed, but the problem is it eventually will look less than ideal when the average display is UHD+...

Edit: as a sidenote, and on topic, does that mean CompA and CompB were rendering out at different debayer settings?

 

[Joe Riggs]

2min vs 35min, is open cl that much faster than cuda?

 

[Christopher-lee dos Santo]

It's perceived sharpness; there is actually more moire/aliasing in the 1/2good render, but when scaled down/back to 1080, it actually looks sharper even though there are less pixels. Again, micro-contrast... Look a the writing on the dry-erase board; it looks "clearer" because there are less intermediate (transition) pixels between the black marker lines are at their thickest, and the whiteboard's white is at its cleanest.

You've seen it before; it's why a camera like the Alexa "looks" just as sharp as 4k/5k, but is only 3.2k. At 1:1, or on a 4k display, or blown up to 70mm projection, you'll see it's shortcomings... On a 1080p monitor, not so much. It's also why a lot of pundits say "4k/resolution doesn't matter"... It's half-true; yes it looks just as good downressed, but the problem is it eventually will look less than ideal when the average display is UHD+...

Edit: as a sidenote, and on topic, does that mean CompA and CompB were rendering out at different debayer settings?

Thanks for the info and this great explanation. That makes perfect sense. And you're right, there is way more aliasing and moire. One can see it in the blinds in the background when the scene plays out. And it's practically gone in the 1/2 premium render. Thank you for this explanation.

It still confuses me as to how the two got exported differently as our rule of thumb was to render out at 1/2 premium for the proxies. In the end, I guess it doesn't really matter as we are relinking to the original R3d 5k files, but I just wanted to find out if there was a difference between the two GPU render options.

Unless one of my assistants rendered out at 1/2 good as opposed to 1/2 premium, which is weird as the settings on both machines for the export preset are identical and are both set to 1/2 premium.

Still confused but it may be possible this is the cause.

Thanks for the help chaps.