madVR - high quality video renderer (GPU assisted)

[huhn]

that's just the file saying it is DCI P3 in bt 2020.
the output and internal colorspace are not shown in the OSD.

[clsid]

New beta 209 of madVR.

[JarrettH]

Wow, I think I'll update to this from 0.92.17. I wasn't into all the 3 month time limits.

[pirlouy]

So he was aware of a problem with HAGS and it could be mitigated by renderer/app ! It's not an isolated problem.
I don't feel like reporting it to affected apps.

[huhn]

there ware massive problems with windows dwm.

[YxP]

HAGS problem seems to be fixed, yay. 10bit colour works for about 30 seconds and then everything gets a sepia filter, but can't complain really, finally something new!

[EDIT]

Looks like Windows 11 tries to apply some sort of LUT to the picture, since activating "override to reference mode" in Nvidia colour settings prevents this from happening. I have no idea if this is a MadVR problem so I'm not going to report a bug.

[clsid]

Report bugs in the AVSForum thread or on Madshi's bug tracker. You can use the debug build of madVR to generate logs.
Now is the time to get bugs fixed. Once Madshi puts his attention back on Envy, it might take another year for a new madVR update.

[strumf666]

Not much difference, if any, with 10bit output on win11/amd graphics card; usually a restart is required if any apps have been opened before movie playback.

[huhn]

this is supposed to be fixed and i never saw that on amd.

edit: no my original report is with AMD. what so ever you should report that.

[JarrettH]

maybe madshi will catch up with the last 1k pages when he left here

[SharpHD]

I do not know anything at all about scripting and would appreciate some help here. Im using my laptop with RTX 3060 with my Samsung Q90R 4K 2019. I would like to automate when madVR plays anything Nvidia should switch from PC mode 4K120Hz 10bit 4:2:0 to 4K60Hz 10bit Full RGB. Is this possible? If yes can it be used with this?

if (srcWidth > 1920) and (srcHeight > 1080) "2160p"

else if (srcWidth > 1280) and (srcWidth <= 1920) "1080p"

else if (srcWidth > 960) and (srcWidth <= 1280) "720p"

else if (srcWidth <= 960) and (srcHeight <= 540) "540p"

[huhn]

well one answer has to be enough.

[Sunspark]

I moved my PC to the old 1366x768 32" PVA panel recently acquired TV which runs at 1360x768 on a PC. I wanted to watch some shows on a larger screen in a softer chair. It's not bad since it has 1:1 pixel, chroma 4:4:4 and full range blacks at 1360. I wish the backlight had a higher PWM rate (phone can see the flashing when you drop the backlight below 90%), maybe I will get used to it, even at a 100% backlight setting it still seems to have a shimmer. Testing displays is a must. This thing absolutely despises RGB input, it bands. You need to use YCbCr. Curiously, if you feed it a 1920x1080 input it doesn't care if it's RGB and the scaler chip will use it without banding or hotspots but overall it probably still prefers YCbCr.

What's been interesting to notice is that on a screen like this you can really see the difference between the various downscaling algorithm choices. On a monitor that is already normal sized or small, you don't really notice that much, but on a larger screen with larger pixels, it's a lot more visible! Leaving chroma upscaling the same, you can see large differences changing the algorithm when downscaling the image from 1920x1080 to 1360x768. Some are sharper, some are smoother, some are grainier, colours are affected (shadows, sunbeams), etc. I'm somewhat surprised how noticeable it is.

[Sunspark]

Saw this post on another forum from a decade ago.

"AMD's DXVA2 algorithms are the most primitive one, Blinear, for all three. NVIDIA image upscaling algorithm differs depending on the source content's resolutions: for SD, it's Lanczos with some edge sharpening filter; for HD and full HD contents, it's Bilinear with with some edge sharpening filter. AMD's and NVIDIA's Chroma Bilinear upscaling is slightly different from madVR's.

Intel's are the best among the three. Chroma upscaling is close to Bicubic 50, image upscaling is Lanczos+AR, image downscaling is a bit difficult to identify, it looks more like Catmull-Rom+AR+unsharp mask than Catmull-Rom+AR+LL. If you are going to use Intel HD Graphics (in any SNB / IVB desktop CPU), madVR DXVA2 is the best choice. When there is no image scaling (mainly full HD contents in a full HD display), you have to choose an chroma upscaling algorithm and it depends on the number of EUs (Execution Units)."

I found this interesting because in part, I saw a comment from huhn awhile ago where he said he doesn't like bicubic150 for downscaling because of motion artifacts or something and prefers catmull-rom/bicubic 50. The dxva downscaling in the older Intel generations (mine is Broadwell) is not to be slept on.. it does appear equally as good or better than a number of the software algos for downscaling.

The info is still relevant for me since this is old hardware I am working with. I wonder if in the years since that was written if AMD or Nvidia have improved things, or if it's still the same as it was back then?

[huhn]

no it has not improved and intel has also utterly changed.

BTW. the old intel kernel is utterly broken and that's why they are the worst.
madVR bicubic beats all of them and it is really not good. DXVA has as said before a broken kernel matric tons of banding issues with chroma position color error and so much more it is the worst. d3d11 api is such an upgrade.

and while it is true that i do not like bicubic 150 in motion and think it should be fixed jinc and ssim are save and better then bicubic. bicbubic 150 is save for upscaled content at least as far as i know there are just other scaler more worth testing.

[Sunspark]

DXVA has as said before a broken kernel matric tons of banding issues with chroma position color error and so much more it is the worst.

This might be generation-specific. I just don't have the chroma issue.. we looked at a test image awhile ago and the shift is not present.

The downside of using that on this system is that yes, while it gives the system a big break on system load, it also means Intel had to be conservative with settings so it wouldn't break on certain content. So it is just average at everything and does have more grain.

It wouldn't surprise me if it was broken today, since everything from the big corporations is vibe-coded slop now.

and while it is true that i do not like bicubic 150 in motion and think it should be fixed jinc and ssim are save and better then bicubic. bicbubic 150 is save for upscaled content at least as far as i know there are just other scaler more worth testing.

I agree. I tested again and decided while it's OK, it could be better. Ssim and jinc provide good visual results but I can't use them right now due to thermal runaway causing the gpu load to ramp up to 100% due to throttling. Waiting for some putty to arrive and then I will replace the fan and all compounds/pads and see if the same thing happens. It's way overdue.

I read somewhere that spline36 (called spline 3-tap in madvr) is a good algorithm to use for non-integer factors when downscaling. When comparing it against lanczos, I agree. Lanczos is subtly sharper, but the problem is you don't want the output to be too sharp because then the eye catches on the edges/grain so spline is better here.

So it's a variable worth considering.. is any given algorithm recommendation better for integer or non-integer scale factors?

4K to 1080p is an integer but 1080p to 768p is not.

[huhn]

ingetar doesn't matter. madVR has some optimization that make it faster for integer but that's it.

yes people like talking about spline doesn't chance that it isn't impressive what so ever since we got AA.

it is irrelevant to talk about the sharpness difference of spline and lanczos when bicubic 150 is on the table a scaler magnitudes sharper then spline and lanczos 3 will ever be. the problem of lanczos was ringing which ahs been fixed.

DXVA scaler are usually utterly free on AMD and nvidia both can be done at idle and the lowerest power state. 480 to 8k or 8k to 480 it doesn't matter.

on older intel cards even the decoder puts load on the GPU which is just sad...

[clsid]

New beta 210 of madVR.

[Schwartz]

Cheers. So all the betas expire in 2027? How about the last stable, does that have a time bomb in it too? Just wondering.

[dbcooper]

Also from Madshi at that forum:

Some good news from Nvidia. I reported the HAGS issue to them with detailed instructions how to reproduce it, and they just told me they were able to reproduce it and will have a deeper look at it.

https://www.avsforum.com/threads/imp...#post-64463606