madVR - high quality video renderer (GPU assisted)

[tobindac]

Even on a forced normal size (view->video frame of mpc), the system appears to be using upscaling (or downscaling) algorithms since by setting all algorithms to Nearest Neighbor I had a significant drop in resources usage.

I was trying to lower system resources when on 1080p video on a 1080p screen.

On full screen.

[huhn]

Even on a forced normal size (view->video frame of mpc), the system appears to be using upscaling (or downscaling) algorithms since by setting all algorithms to Nearest Neighbor I had a significant drop in resources usage.

I was trying to lower system resources when on 1080p video on a 1080p screen.

On full screen.
even with 1080p on a 1080p screen chroma upscaling is used because nearly every source today is 4:2:0
so the 540p chroma has to be scaled to 1080p

[tobindac]

even with 1080p on a 1080p screen chroma upscaling is used because nearly every source today is 4:2:0
so the 540p chroma has to be scaled to 1080p

Ah nice. Good to know since on old hardware I may want to actually avoid a slow algorithm if the hardware can't handle it.

[huhn]

Ah nice. Good to know since on old hardware I may want to actually avoid a slow algorithm if the hardware can't handle it.

use bilinear it is not really slower than nearest neighbor but a lot better.

you can use some of the trade quality for performance options to like:
"use 10 bit chroma buffer instead of 16 bit"
"use 10 bit image buffer instead of 16 bit"

if you can avoid nearest neighbor with these options do it.

[leeperry]

So did anybody make real-world comparisons proving that 64 neurons luma 720@1080p looks noticeably better than 32 with movies? Just to keep placebo at bay :)

A local online retailer is currently giving a 50% MIR on the TF4 MSI 750Ti & 270, the former is barely as fast as the 7850 but is highly power efficient....OTOH, the 270 is a rebadged 7870 and is quite a bit faster than even my o/c 7850 so I could even go 64x luma on 29.97 720p.

Of course the 750Ti at full load needed 141W in a local comparison I found, my o/c 7850 214W and the 270 244W. Sucking up 250W to watch movies doesn't sound too reasonable, OTOH I'm not a hardcore gamer and I really only watch 1 movie/day on an average so cry me a river :D

[michkrol]

So did anybody make real-world comparisons proving that 64 neurons luma 720@1080p looks noticeably better than 32 with movies? Just to keep placebo at bay :)
To be honest, I can't differentiate between 720p@1080p with Jinc3AR upscaling and NNEDI 32/64 neurons + Catmull-Rom downscaling with most of content I've used for testing, even while pausing, hence I'd recommend the GTX 750Ti ;)

[tobindac]

use bilinear it is not really slower than nearest neighbor but a lot better.

you can use some of the trade quality for performance options to like:
"use 10 bit chroma buffer instead of 16 bit"
"use 10 bit image buffer instead of 16 bit"

if you can avoid nearest neighbor with these options do it.

Thanks. I currently figured Softcubic is very light compared to a huge impact with Lanczos X3. Weird, it may be a chroma-only thing.

[3ll3d00d]

(Sorry if this has been answered before but I can't find an answer)

how do image doubling and upscaling interact? more specifically can you control which one is used? also, does it choose an algorithm for each dimension individually or for the source as a whole?

For example, if you have PAL DVD & a 1080p screen then the stats report 720x576 as the source and 1920x1080 as the output. Assuming image doubling is set to "always", does it double to 1440x1152 and then upscale horizontally and downscale vertically? or does it just upscale only?

[Asmodian]

how do image doubling and upscaling interact? more specifically can you control which one is used? also, does it choose an algorithm for each dimension individually or for the source as a whole?

You cannot control which is used explicitly but you can implicitly. If the criteria is such that image doubling is used the image is doubled in both dimensions (scaling factor > 1.2, etc.). Upscaling or downscaling is then used to bring the image to display dimensions as needed. Note Jinc cannot be for downscaling so Lanczos is used for both up and downscaling if upscaling is set to Jinc and you need to upscale in one dimension and downscale in the other.

For example, if you have PAL DVD & a 1080p screen then the stats report 720x576 as the source and 1920x1080 as the output. Assuming image doubling is set to "always", does it double to 1440x1152 and then upscale horizontally and downscale vertically? or does it just upscale only?

Assuming square pixels the image would be doubled to 1440x1152 and then downscaled in both dimensions to 1350x1080. If anamorphic it would be upscaled horizontally and downscaled vertically after being doubled.

[huhn]

Thanks. I currently figured Softcubic is very light compared to a huge impact with Lanczos X3. Weird, it may be a chroma-only thing.

with a 1080p source at 1080p it should be chroma only.

how do image doubling and upscaling interact? more specifically can you control which one is used? also, does it choose an algorithm for each dimension individually or for the source as a whole?
image doubling always uses NNEDI3, Asmodian said the rest.

[chros]

@Madshi: wooow! Thanks for the profile support!!!

Now I can enjoy the regular content on my reduced powered m650GT laptop in top quality (chroma and luma upscaling Jinc+AR) and the rest (mainly HD sport) in lower quality (chroma and luma upscaling Lanczos+AR) with this:

if (srcWidth >= 1280) and (srcFps >= 29) "HD60fps"
else "Default"

[leeperry]

To be honest, I can't differentiate between 720p@1080p with Jinc3AR upscaling and NNEDI 32/64 neurons + Catmull-Rom downscaling with most of content I've used for testing, even while pausing, hence I'd recommend the GTX 750Ti ;)
I'm sitting 80cm away from a 32" and I think I can see differences when really really looking for them, this said when sitting on the couch from a 2 meters distance that's another matter.

IME J3AR looks blurry, 16x luma is a huge improvement, 32x doesn't seem to look as sharp as 16x but 64x looks sharper, less noisy and more genuine/natural than 16x.....the PQ improvement isn't really worth an extra 100W, though.

All this said, it's a known fact that in the A/V hobby the last few percents will cost ya. I'll run more tests today before pulling the trigger on either of them :p

[ryrynz]

and I think I can see differences

LOL. :p


I'll run more tests today before pulling the trigger


Yeah, I swear you got a thing for pulling triggers..

[StinDaWg]

There's a big difference between Jinc and NNEDI on my 50" tv. NNEDI is sharper and text looks basically perfect and isn't aliased like all the other upscalers. There's also less ringing.

[leeperry]

LOL. :p

Yeah, I swear you got a thing for pulling triggers..
Well, a 50% MIR on a dead-silent TF4 MSI is hard to refuse :)

Differences aren't quite night & day either way, but 64x NNEDI does appear to look the clearest..whether it's worth the extra power consumption is another story...OTOH faster Maxwell's will be out by the end of the year so it's not like I'd keep the 270 for very long.

[3ll3d00d]

You cannot control which is used explicitly but you can implicitly. If the criteria is such that image doubling is used the image is doubled in both dimensions (scaling factor > 1.2, etc.). Upscaling or downscaling is then used to bring the image to display dimensions as needed. Note Jinc cannot be for downscaling so Lanczos is used for both up and downscaling if upscaling is set to Jinc and you need to upscale in one dimension and downscale in the other.



Assuming square pixels the image would be doubled to 1440x1152 and then downscaled in both dimensions to 1350x1080. If anamorphic it would be upscaled horizontally and downscaled vertically after being doubled.

Ok thanks. So basically it doubles first if configured to do so and then scales to fit.

[fairchild]

16x luma is a huge improvement, 32x doesn't seem to look as sharp as 16x but 64x looks sharper, less noisy and more genuine/natural than 16x.....the PQ improvement isn't really worth an extra 100W, though.

16x less sharp than 32x? Seems strange unless I'm missing something, seeing how madshi made 32x the default. I have been using 32 and 64 neurons using this profile:

if (srcHeight <= 480) "480p"
if (srcHeight <= 720) "720p"
else "1080p"

480p: NNEDI3 64 neurons (2x) ~~doubling actives on 960x540 or less~~
720p: NNEDI3 32 neurons (2x) ~~doubling actives on 960x540 or less~~

I don't use 1.5x since I don't see a difference with upscaling 720p to 1080p. (even with CR AR+LL on downscaling, feel that it's unnecessary to double over display resolution only to then downscale, added step and all)

I use the following scalers:

chroma upscaling lanczos3 (AR)
image upscaling: lanczos3 (AR)

[Asmodian]

Ok thanks. So basically it doubles first if configured to do so and then scales to fit.

That is a much more concise way to put it, yes. :)

[GCRaistlin]

no details lost when i watch the "dark" picture with a 3dlut and madvr.

what type of gpu are you using. looks like a this is a simple mismatch with hdmi tv settings and gpu settings.

most likely madvr does nothing wrong.

can you try these samples to make sure? https://www.wuala.com/alluringreality/Public/MP4-2c.7z
source: http://www.avsforum.com/t/948496/avs-hd-709-blu-ray-mp4-calibration

non calibrated display normally always clip black!

I use GF8800GT.

Finally I have found time to try these samples. I have calibrated my monitor before using other tool, CLTest, but these samples... I really don't understand what is going on but, according to the manual, I should set brightness to 0 'cause even by zero brightness I can see reference black and below! But I have gray color instead of white by zero brightness...

Anyway, how can my monitor settings affect the way madVR displays the video? Screenshots in my post (http://forum.doom9.org/showthread.php?p=1664958#post1664958) look different, don't they? One of them will always be darker than another one, regardless of monitor settings, right? So how can I set madVR to produce the picture as light as VMR9 to not lose the details?

[huhn]

I use GF8800GT.

Finally I have found time to try these samples. I have calibrated my monitor before using other tool, CLTest, but these samples... I really don't understand what is going on but, according to the manual, I should set brightness to 0 'cause even by zero brightness I can see reference black and below! But I have gray color instead of white by zero brightness...

Anyway, how can my monitor settings affect the way madVR displays the video? Screenshots in my post (http://forum.doom9.org/showthread.php?p=1664958#post1664958) look different, don't they? One of them will always be darker than another one, regardless of monitor settings, right? So how can I set madVR to produce the picture as light as VMR9 to not lose the details?


get this file: http://www.w6rz.net/avshd709/MP4-2c.exe
source: http://www.avsforum.com/t/948496/avs-hd-709-blu-ray-mp4-calibration

open the 1- black clipping file and report back!

test it in both VMR and MadVR. after this post your MadVR settings from device -> "monitor name" -> properties.

[JohnLai]

madVR does not support ICC profiles, though it will be affected by any gamma ramps that the ICC profile loads into your GPU. To get the most out of your profile, you should make a Rec.709 3DLUT out of it using ArgyllCMS (coupled with dispcalGUI if you don't like console commands). For more instructions see this thread (http://www.avsforum.com/t/1471169/madvr-argyllcms) (though the first post is a bit out of date). You can use ArgyllCMS for calibration and profiling as well, but your existing profile should work fine for making a 3DLUT.

The 'disable GPU gamma ramps' option is for when your 3DLUT already includes the calibration, either integrated into the values or appended. This will disable the calibration loaded into your GPU, so other programs will look different while madVR is active. The exception to this is when you use the windowed overlay mode, but I forget what the right settings are for that :)

Oops, sorry for my late reply, being busy with life T_T

I tried to use ArgyllCMS with displacalGUI, but it seems to be too complicated for my level.

How about if I use ICC Profile Inspector to find the value of XYZ (red,green,blue,white point) from the ICC profile created by Spyder4? These are the value ;

red, XYZ, 0.39713, 0.23576, 0.05164
green, XYZ, 0.35269, 0.59932, 0.11452
blue, XYZ, 0.21439, 0.16492, 0.65874
white, XYZ, 0.96420, 1.00000, 0.82491

Btw, these are the value for my laptop LCD.

Can I just select 'calibrate this display by using yCMS' and enter the above value to 'primaries / gamut measurements'?

Since I can't get the value for 'grayscale / gamma measurements', I will leave it empty. For compensation of this, the 'disable GPU gamma ramps' will not be checked since Spyder4 Utility already applies the calibration (color and gamma adjustment) systemwide.

[mzso]

Hello!
Are there any DVB(-C) software out there that can use madVR as a renderer? Potplayer which I use for videos claims to support BDA drivers, but it doesn't actually work with my DVB-C tuner.

[StinDaWg]

I don't use 1.5x since I don't see a difference with upscaling 720p to 1080p. (even with CR AR+LL on downscaling, feel that it's unnecessary to double over display resolution only to then downscale, added step and all)
Interesting. We have differing opinions. I see a large sharpness increase with NNEDI, less scaling artifacts, and sharper text without the aliasing. Try Lanczos4 for downscaling instead of CR, if you still don't see the difference I don't know what to tell you. How large is your tv and how far do you sit away from it?

Take a look at the ticker, there is a huge difference between Lanczos3 and NNEDI3+Lanczos4 or NNEDI3+CR
http://s000.tinyupload.com/index.php?file_id=64213340499639971921

[QBhd]

Hello!
Are there any DVB(-C) software out there that can use madVR as a renderer? Potplayer which I use for videos claims to support BDA drivers, but it doesn't actually work with my DVB-C tuner.

have you been here:
http://www.dvbsupport.net/

I don't use DVB, but it may have some useful info for you

and I think this may be a little Off Topic...

QB

[fairchild]

Interesting. We have differing opinions. I see a large sharpness increase with NNEDI, less scaling artifacts, and sharper text without the aliasing. Try Lanczos4 for downscaling instead of CR, if you still don't see the difference I don't know what to tell you. How large is your tv and how far do you sit away from it?

Take a look at the ticker, there is a huge difference between Lanczos3 and NNEDI3+Lanczos4 or NNEDI3+CR
http://s000.tinyupload.com/index.php?file_id=64213340499639971921

I have a 32" LCD TV (sony 32EX400) which I sit around 2 feet from, and a 55" plasma TV (panasonic 55VT60) which I sit around 6-10 feet.

What am I looking for in that clip, a particular banner letter? Because I did a quick comparison and don't really see a difference unless I'm staring/looking for a particular thing, in which case not sure it's going to be noticeable with most video content. (non-ESPN)

[leeperry]

16x sharper than 32x? Seems strange
Been running more tests with CR AR LL downscaling, J3AR chroma & mono-static A4 LL on very crisp & colorful REC.709 high-motion 720p25 content and to my eyes 16x looks sharper than 32x, 32x looks more detailed than 16x but softer and it's like 64x comes with the details of 32x and the sharpness of 16x....that's a one-way for me, it looks darn close to native 1080p...we'll see a few years from now when we'll be able to go 256x :sly:

If someone made a NNEDI ASIC, he would take the post-production/videophile market by storm IMHO. Darbee be gone ^^

Still need to scrutinize what looks best on SD but any mid-end GPU can do 256x so there's no hurry, I would also need to figure out what to use for chroma as I've grown the habit of mixing softish chroma & sharp luma and I prefer to use GPU cycles to warrant 64x luma on 720p. NNEDI on SD chroma would appear to make artifacts more visible IME, that's counterproductive.

[vBm]

Hello!
Are there any DVB(-C) software out there that can use madVR as a renderer? Potplayer which I use for videos claims to support BDA drivers, but it doesn't actually work with my DVB-C tuner.

You could try MPC-HC ... and if it doesn't work with your DVB-C tuner, kasper93, Underground78 or XPC1000 could make it work, just remember to open a ticket on trac.mpc-hc.org

[JohnLai]

Oops, sorry for my late reply, being busy with life T_T

I tried to use ArgyllCMS with displacalGUI, but it seems to be too complicated for my level.

How about if I use ICC Profile Inspector to find the value of XYZ (red,green,blue,white point) from the ICC profile created by Spyder4? These are the value ;

red, XYZ, 0.39713, 0.23576, 0.05164
green, XYZ, 0.35269, 0.59932, 0.11452
blue, XYZ, 0.21439, 0.16492, 0.65874
white, XYZ, 0.96420, 1.00000, 0.82491

Btw, these are the value for my laptop LCD.

Can I just select 'calibrate this display by using yCMS' and enter the above value to 'primaries / gamut measurements'?

Since I can't get the value for 'grayscale / gamma measurements', I will leave it empty. For compensation of this, the 'disable GPU gamma ramps' will not be checked since Spyder4 Utility already applies the calibration (color and gamma adjustment) systemwide.

Hmm, I noticed Argyll_V1.6.3 can only load ICCv2, and so I recalibrate again with a new ICC/ICM v2 instead of v4 file created by Spyder4Elite 4.5.4

Next, I ran dispcalGUI-3DLUT-maker.exe directly trying to create madvr 3DLUT. But I encountered issues.

If I check the box 'Apply Calibration (vgct)' and got it loaded by madvr, I got black screen during playback.

However, if I uncheck that particular box, madvr playback is fine, but I am not sure if this is the correct way. Most guide found online recommended to check that box.

I attached my ICCv2 profile generated by Spyder4Elite software.
Anyone can figure out what went wrong?:p

[StinDaWg]

I have a 32" LCD TV (sony 32EX400) which I sit around 2 feet from, and a 55" plasma TV (panasonic 55VT60) which I sit around 6-10 feet.

What am I looking for in that clip, a particular banner letter? Because I did a quick comparison and don't really see a difference unless I'm staring/looking for a particular thing, in which case not sure it's going to be noticeable with most video content. (non-ESPN)
Letters that have diagonal lines (M, N, A, W, V, ect) look jagged (aliased) with any of the upscaling algorithms besides NNEDI3. There is also less ringing in the whole picture compared to straight up Lanczos3 upscaling.

Been running more tests with CR AR LL downscaling, J3AR chroma & mono-static A4 LL on very crisp & colorful REC.709 high-motion 720p25 content and to my eyes 16x looks sharper than 32x, 32x looks more detailed than 16x but softer and it's like 64x comes with the details of 32x and the sharpness of 16x....that's a one-way for me, it looks darn close to native 1080p...we'll see a few years from now when we'll be able to go 256x :sly:
I'm having a hard time understanding what you mean. If 32 looks softer, then it should look less detailed as well, not more detailed. madshi said the more neurons you use the better the quality gets, so I find it hard to believe that 32 looks worse than 16. Maybe he can chime in, but I haven't seen him post in here for awhile.

NNEDI3 is an excellent upscaling algorithm, but my eyes haven't been tricked into thinking it was native 1080p. I can still tell the difference.

[leeperry]

If 32 looks softer, then it should look less detailed
Details and Sharpness are two different things, anyway point of my story is that 64x on 720p@1080p looks fantastic...maybe it's got to do with the very crisp 720p25 video material I've been using that's far cleaner than a 24p movie. Not going back to 32x land :p

[Devrim]

Is it possible to stop madvr using GPU when the video is paused? When you pause the video madvr still seems to be using/triggering the GPU.

[StinDaWg]

Details and Sharpness are two different things.
Yes and no, but they are related. I don't really feel like typing up a long response so just watch whatever you enjoy. :)

[leeperry]

When you pause the video madvr still seems to be using/triggering the GPU.
Yeah, now when I want to pause a movie for a few minutes I simply close my media player...don't see the point of wasting electricity.

Apparently some GPU's can shutdown completely when the display is turned off, but that doesn't seem available for desktop systems. (http://forums.anandtech.com/showthread.php?t=2186725)

[Devrim]

Yeah, now when I want to pause a movie for a few minutes I simply close my media player...don't see the point of wasting electricity.

Apparently some GPU's can shutdown completely when the display is turned off, but that doesn't seem available for desktop systems. (http://forums.anandtech.com/showthread.php?t=2186725)

My GPU clocks back when it is not used but it stays at normal speeds when a video is open.

[Meulen92]

First off all I want to thank madshi for this beautiful renderer. Ive been using it for the last two years.

I have a few questions for you guys on this forum. I hope you could answer them for me:

- What would be prefered next to 256 neuron Luma doubling: 64 neuron Chroma doubling or 64 neuron Luma quadrupling?

- At what number of neurons will NNEDI generally beat Jinc3AR in chroma upscaling?

Thank you in advance!
And madshi, keep up the good work!

[michkrol]

- What would be prefered next to 256 neuron Luma doubling: 64 neuron Chroma doubling or 64 neuron Luma quadrupling?
It depends on source and target resolution or rather scaling factor. If not scaling more than 2x, quadrupling is just pointless. Other than that, it's personal preference and spotting or not spotting the difference depends on your display and viewing conditions.
Those settings are just insane, by the way. Not sure what setup you have (or how low resolution your source is?) to make it smooth.
- At what number of neurons will NNEDI generally beat Jinc3AR in chroma upscaling?
That's hard to answer, the results will vary greatly based on source material and personal preference comes into play here, so you need to try it for yourself. Chances are you won't spot any difference.

[Meulen92]

Thanks for your fast response and the great advice!

Im trying to upscale some 360p material to 1080p using an overclocked HD7950 and a 3770K at 4.8 GHz. So far im able to chroma scale with 256 NNEDI, upscale with Jinc3AR and double Luma with 256 neurons, without dropping frames. Theres still some performance room left for another enhancement.

[leeperry]

Im trying to upscale some 360p material to 1080p using an overclocked HD7950 and a 3770K at 4.8 GHz. So far im able to chroma scale with 256 NNEDI, upscale with Jinc3AR and double Luma with 256 neurons, without dropping frames. Theres still some performance room left for another enhancement.
A 7950 will already make your PC suck more or less 350W at full load, you need to decide whether the stellar power consumption is worth the PQ improvement(if any) and I still feel that on 360p NNEDI will make chroma artifacts more visible than J3AR.

[Meulen92]

Thanks for your feedback! Power is not a concern. Im also a heavy gamer and im on a cfx setup. So while playing BF4 im probally peaking at ~900W at times.
Ill try your sugestion and ill report back tomorrow.
Thanks!

[leeperry]

900W, crazy :D

Very often I find NNEDI too sharp for chroma, even on untouched 1080p sources...to the point that I start wondering if H264 encoders don't sharpen up chroma in order to overcome the details loss due to the half-resolution compression.

But that's on a 4:2:2 TV, maybe its internal processing is already sharp to begin with(even though I turned off its sharpness enhancement setting), OTOH my brain doesn't appear to like chroma noise so YMMV. J3AR is the perfect balance IME, it looks wonderful and very natural with >=64 neurons for luma :)

[Andy o]

Is it possible to stop madvr using GPU when the video is paused? When you pause the video madvr still seems to be using/triggering the GPU.

Are you using DXVA or any other hardware decoding?

[Meulen92]

900W, crazy :D

Very often I find NNEDI too sharp for chroma, even on untouched 1080p sources...to the point that I start wondering if H264 encoders don't sharpen up chroma in order to overcome the details loss due to the half-resolution compression.

But that's on a 4:2:2 TV, maybe its internal processing is already sharp to begin with(even though I turned off its sharpness enhancement setting), OTOH my brain doesn't appear to like chroma noise so YMMV. J3AR is the perfect balance IME, it looks wonderful and very natural with >=64 neurons for luma :)


Thanks for your input!

I wonder how other people feel about using NNEDI for chroma upscaling. Someone who'd like to share his/her opininion?

As for doubling, leeperry, do you think that doubling chroma makes any sense when using J3AR for chroma upscaling? For example when upscaling 320p.

[iSunrise]

VESA finally approved adaptive-sync for the masses!

VESA® Adds ‘Adaptive-Sync’ to Popular DisplayPort™ Video Standard (http://www.vesa.org/featured-articles/vesa-adds-adaptive-sync-to-popular-displayport-video-standard/)

While this wouldn't make Reclock completely obsolete, this is a major step we have all been waiting for, not just for use with games, but also for madVR and jutter-free video playback without having to worry (hopefully) about custom refresh rate tweaking.

[Procrastinating]

While custom refresh rate tweaking wasn't really an issue, it does make everything a lot more convenient. A very nice addition indeed.

[nevcairiel]

Now only more than weird TN gamer screens need to adopt such features.

[StinDaWg]

I wonder how other people feel about using NNEDI for chroma upscaling. Someone who'd like to share his/her opininion?
I don't see any difference between bicubic and nnedi3 with chroma on my Panasonic plasma. Image upscaling by far makes the most noticeable difference.

[cyberbeing]

Use of NNEDI3 Luma Doubling & Chroma Upscaling together can make a rather significant difference on SD content and below. Chroma Doubling is much tougher sell considering its high cost. The majority of the time, you are better off finding your highest usable Luma Doubling setting, and only then considering use of NNEDI3 chroma if it can fit easily within the remaining GPU headroom.

[Meulen92]

Ive followed up your advice and Ive made profiles for all resolutions. I must say pq has seriously improved. So, thanks for your post!

[Devrim]

Are you using DXVA or any other hardware decoding?

Yes,

I'm using CUVID (LAV Filters)

edit

according to nevcairiel it is a CUVID problem. It is just the way it works.

[Devrim]

Anyone else experiencing frame drops reported by mpc-hc when turning on smooth motion in fullscreen?

The video looks smooth to me but mpc-hc reports about 140 frames dropped on 2000 frames drawn.

Also the framerate seems to be 44ish on a 23,976fps video