madVR - high quality video renderer (GPU assisted)

[huhn]

Originally Posted by nevcairiel

I can only repeat what various people have told you before:

29i is supposed to be output as 59p. (Note that many people mix 29i and 59i, but generally mean the same thing, 29 interlaced frames/59 fields per second, ie. what is typically broadcast in NTSC countries)
If its interlaced, it should be output with twice the "frame rate". Its as simple as that.

That sample however is NOT 29i, its 29p. You renamed it with a "deint=video" tag in the filename, which FORCES madVR to deinterlace the video. If you do that, its going to treat it like a normal interlaced file, which as a result does indeed duplicate all frames - but thats only because you forced it to do that, it wouldn't do that on its own.

Interlaced video does NOT contain frames, it contains fields. For technical reasons, two fields are packed into one "frame". So what you do when deinterlacing is first split these two fields again - so your 29 fps interlaced stream becomes 59 fields per second, and then every field is individually deinterlaced, which results in 59 frames per second. Thats how proper deinterlacing HAS to work. If a stream is interlaced, that is what will ALWAYS happen.

i'm pretty sure his main quetion is about 29i with telecine 2:2 content if it is fine to watch this with half frame rate.

a 1080p29 interlaced stream with telecine is normally called telecine in this forum to make the difference clear but the stream is still interlaced.

of cause this doesn't change all the fakt you have stated above proper deinterlacing creates 50/60 fps even with a 3:2 telecine source. a very clear high quality BD with 1080p29i and 2:2 cadence is deinterlanced to 59p with duplicated frames thanks to the inverse telecine function of nvidia and thanks to the pulldown detection of amd.
but in the end a stream like this should be watched with MadVR IVTC or in very rare cases just as progressive.

[Procrastinating]

I figured that it might help madshi and those who want to keep track of the thread if a survey was created which allowed people to enter render times, settings and GPU in a formatted manner.

To achieve that, I created a public survey. The spreadsheet is linked upon submission of some render times. If anyone has questions, suggestions or problems with the survey, feel free.

I chose a default setting which is probably a realistic benchmark for everyday use of NNEDI3.

[ryrynz]

Madshi, how does madVR's NNEDI settings stack up vs the default settings in Avisynth filter? Was wondering if we'd see options available for nsize and qual etc.

[nevcairiel]

Originally Posted by Procrastinating

I chose a default setting which is probably a realistic benchmark for everyday use of NNEDI3.

Using NNEDI for chroma upscaling is rather wasted, as the benefit is rather miniscule. This does not seem to be a realistic setting.
Luma doubling is its real power and its real visible advantage over the other modes.

Also, madVR cannot make use of multiple GPUs, mentioning the whole CF/SLI stuff just overcomplicates it.

[ryrynz]

Originally Posted by Procrastinating

I figured that it might help madshi and those who want to keep track of the thread if a survey was created which allowed people to enter render times, settings and GPU in a formatted manner.

Is it possible to use drop down boxes for options? People can just test what combinations they want which are likely to be what others will want to run anyway.

[jkauff]

Originally Posted by nevcairiel

Using NNEDI for chroma upscaling is rather wasted, as the benefit is rather miniscule. This does not seem to be a realistic setting.

I think you mean chroma doubling, right?

[madshi]

Originally Posted by Procrastinating

I figured that it might help madshi and those who want to keep track of the thread if a survey was created which allowed people to enter render times, settings and GPU in a formatted manner.

To achieve that, I created a public survey. The spreadsheet is linked upon submission of some render times. If anyone has questions, suggestions or problems with the survey, feel free.

I chose a default setting which is probably a realistic benchmark for everyday use of NNEDI3.

This is interesting, but I would suggest to wait for the next madVR build before doing the survey, because I'm currently trying to reduce the AMD OpenCL interop cost.

[huhn]

you find a way to lower data size, so NN is performed on the opencl side?
or are you going with something like d3d -> direct compute -> opencl?
time will tell

[nevcairiel]

Originally Posted by jkauff

I think you mean chroma doubling, right?

His one pre-defined settings string suggested NNEDI for chroma upscaling, and thats what I meant.

Chroma doubling is probably in a similar boat, though.

[Procrastinating]

Originally Posted by madshi

This is interesting, but I would suggest to wait for the next madVR build before doing the survey, because I'm currently trying to reduce the AMD OpenCL interop cost.

I figured I would let people select any particular build they want at this stage, any data is good data. If you want only a subset of information in the survey though, then I can change that.

Originally Posted by ryrynz

Is it possible to use drop down boxes for options? People can just test what combinations they want which are likely to be what others will want to run anyway.

Google docs seems to be impressively inflexible, but I'm sure I could try and find a way to turn all the convoluted typing into a series of drop-down boxes, which I think everyone would prefer. I need sleep though.

Originally Posted by nevcairiel

Using NNEDI for chroma upscaling is rather wasted, as the benefit is rather miniscule. This does not seem to be a realistic setting.
Luma doubling is its real power and its real visible advantage over the other modes.

Also, madVR cannot make use of multiple GPUs, mentioning the whole CF/SLI stuff just overcomplicates it.

I have heard a few people find that chroma upscaling actually benefits more than luma with NNEDI3 in some cases, but if the semi-general consensus is something entirely different from what I understand, then I can change that.

I included crossfire because madshi said that while he isn't trying to work on it, he is not trying to prevent it either. I think it would be interesting nonetheless to see if crossfire performance improves or degrades over time, for when/if madshi wants to start working on crossfire.

The main draw for this I think, besides providing more efficient data for madshi, would be to track performance in various scenarios and hardware configs over time. If I can sort out a nice way to both input and parse information quickly and painlessly with this, it could become quite a handy tool for everyone. I'm thinking about seeing if I can parse specific GPUs and compare their characteristics against rendering times with this system. Of course, the spreadsheet is free for anyone to make use of as well.

[DragonQ]

Originally Posted by madshi

This is interesting, but I would suggest to wait for the next madVR build before doing the survey, because I'm currently trying to reduce the AMD OpenCL interop cost.

I suppose I will be your prime guinea pig for that?

[michkrol]

Originally Posted by Procrastinating

I figured that it might help madshi and those who want to keep track of the thread if a survey was created which allowed people to enter render times, settings and GPU in a formatted manner.

To achieve that, I created a public survey. The spreadsheet is linked upon submission of some render times. If anyone has questions, suggestions or problems with the survey, feel free.

I chose a default setting which is probably a realistic benchmark for everyday use of NNEDI3.

Seems like a good idea, however to make it actually comparable you'd need to impose all settings, possibly some predefined presets to chose from and impose the file played. Best to use some freely available material like Big Buck Bunny (www.bigbuckbunny.org), available under Creative Commons rights. That way you can be sure that the testing environment is similar for everyone.
I'd also include GPU usage apart from rendering times, since with NNEDI you get a penalty from just enabling it and increasing neurons doesn't increase the render times that much, especially on stronger GPUs.
Then you just take the results, group them by target (screen's) resolution and settings (presets) to compare them pretty easily.

That's of course just my opinion, but I hope it proves useful for you.

[JonnyRedHed]

Would it be possible to have the OSD display which profile is active when playing please. And to also highlight said active playing profile in the MadVR settings window.

I have to currently look at my GPU usage and temp to work out what's going on. I'm not a very technical user so something as simple as a line on the OSD and a highlight over the profile in settings would be really useful.

[michkrol]

Originally Posted by JonnyRedHed

Would it be possible to have the OSD display which profile is active when playing please. And to also highlight said active playing profile in the MadVR settings window.

I have to currently look at my GPU usage and temp to work out what's going on. I'm not a very technical user so something as simple as a line on the OSD and a highlight over the profile in settings would be really useful.

The active profiles are bolded in settings, sometimes you have to close and re-open the settings for this to work, especially after changing auto-select rules.

Showing active profiles in OSD would be kind of impractical, since you can have multiple profiles groups, hence multiple active profiles at once (up to 9 right now).

[nevcairiel]

Originally Posted by JonnyRedHed

Would it be possible to have the OSD display which profile is active when playing please.

Depending on your settings, you could have a multitude of profiles active, since you can define multiple profile groups for different settings.
So this is rather impractical for the OSD.

Originally Posted by JonnyRedHed

And to also highlight said active playing profile in the MadVR settings window.

The active profile is highlighted in bold right now.

PS:
michkrol was faster

[JonnyRedHed]

OSD, Ah ok I understand.

Bolded in settings, ah yes close and re-open. Thank you for explaining.


I can see them working now, thanks.

[Farfie]

Regardless of whether or not I'm upscaling from 720p to 1440p, NN looks like crap. The image is fuzzy and aliased as hell. I realize the theory behind NN for a 1:1 mapping in this scenario, but it does not work out for me.

In my opinion, for anime especially, NNEDI3 is the obvious champion. Even at only 16 neurons, it increases sharpness drastically while keeping aliasing low, even if it is undoubtedly more aliased than jinc, to my eyes. Still as I said, I believe it keeps the aliasing to a minimum, so only if you find yourself extremely sensitive to aliasing should you bounce back and forth between NNEDI3 and Jinc to determine what you find more pleasing. Spline and Lanczos are still great, but I believe they should be reserved for weaker hardware setups. To me, everything else is deprecated. And again, this has been my opinion

[Ceremony]

Originally Posted by madshi

This is interesting, but I would suggest to wait for the next madVR build before doing the survey, because I'm currently trying to reduce the AMD OpenCL interop cost.

about that test:
I am unable to run it on my A10-7850K. It crashes with the following error instantly:

The application has failed to start because its side-by-side configuration is incorrect. Please see the application event log or use the command-line sxstrace.exe tool for more detail.

do i need to install anything before running the benchmark your provided at http://madshi.net/OpenClSpeedTest.zip

[Fullmetal Encoder]

Originally Posted by Asmodian

Why? What is worse about 480p/576p -> 2160p compared to 480p/576p -> 1080p?

Watching a DVD on a 70" screen I can understand some shudders but that is due to the size not the resolution, the extra resolution only helps.

It would be worse since the larger the screen resolution is compared to the resolution of the source image the larger the number of pixels that must invented to fill that higher resolution in the process of interpolating the image. Generally speaking I think that an image that has thousands of guesses as to what should be in those pixels is less desirable than an image that has hundreds of guesses. Unless, of course, there is something seriously awry with my understanding of how scaling is done.

[John Carmack]

Originally Posted by Ceremony

about that test:
I am unable to run it on my A10-7850K. It crashes with the following error instantly:


do i need to install anything before running the benchmark your provided at http://madshi.net/OpenClSpeedTest.zip

I had this error too. I downloaded an obscure VCredist 2005 from MS and it worked.