madVR - high quality video renderer (GPU assisted)

[secvensor]

I suggest to give the chance to adjust color and a transparency of panels Exclusive mode.

[turbojet]

I spoke too soon about auto play working again with latest madvr+mpc-be, it only works when player is opened on primary display then moved to extended display when it used to pause. If player is open open extended display it starts paused every time still. Madshi can you let me know if it's madvr or mpc-be's issue to fix?

While I agree it really isn't the video renderer's responsibility to fix badly encoded files but it does the job currently, same could be said for colormatrix issues. I found out it's x264 to blame for the damaged lines when height is not mod4 and reported it.

[chros]

Thank you, madshi for implementing the custom pixel shader trade-offs!

On my laptop the result is:
1. doesn't make a difference if it's on or not: store custom shader results in 16bit...
2. the other is making the difference: run custom shaders in video level...
3. "don't use copyback for DXVA decodeing" causes bigger GPU usage if it's enabled !!!
(using LAV dxva native) GPU:62%/69% , VPU: 23%/27% !!! , MCU: 42%/50%
And when I enable it on the fly it results constant green picture (it works fine in the opposite way).

Edit: tested with the v306.97, now I realised that a newer whql is out, I will test them again ...
Edit2: situation is the same with the new v310.70 driver.

[ryrynz]

I honestly don't think my eyes are deceiving me, overlay looks to be smoothest option.
I don't have dropped frames without overlay, possible or placebo? Each time I look at overlay I can see it, like it's rendered to the screen ever so slightly faster.

[annovif]

v0.85.5:

* up to 56% speed improvement for Jinc3 chroma upscaling
* up to 40% speed improvement for Jinc3 AR chroma upscaling

Madshi you are a genius, I was buying a new video card and suddenly my 6670 is powerful enough, but..........What did you done? A pact with the :devil:? Are you sure there is the same quality in this Jinc3 than before? I have a bad feeling about this and jinc3 works too good to be ruined by compromise.

Good year, Fabio

[nevcairiel]

By hard-coding certain parameters, you can easily speed up algorithms, i have no doubt that its still at 100% quality.
Of course these optimizations only work for certain special cases, in this case a clear 2x scaling, which chroma typically is.

[DragonQ]

Surely 2x is rarely used for luma. Chroma 2x is obviously very useful for 1080p content on a 1080p screen but no-one has any 960x540 videos, do they? I guess those with 2560x1440 screens playing 720p content would count but if you have screens that size, surely your content would be mostly 1080p.

In any case, Jinc for chroma with 2x scaling still isn't fast enough on my HD4000 (playing a 1080i/25 file on a 1080p screen).

[sneaker_ger]

Surely 2x is rarely used for luma. Chroma 2x is obviously very useful for 1080p content on a 1080p screen but no-one has any 960x540 videos, do they? I guess those with 2560x1440 screens playing 720p content would count but if you have screens that size, surely your content would be mostly 1080p.

Doesn't madVR first scale chroma to luma resolution, so it would be useful for all 4:2:0 content, regardless of display resolution/video resolution?

[06_taro]

I wonder if upscaling to 2x by Jinc3 and then downscaling to 1.9x by Bicubic (or any other 2-tap stuffs, or even Bilinear) would be faster than directly upscaling to 1.9x by Jinc3 or not since 0.85.5. If it's significantly faster, it might probably worth a try even if there is a tiny quality loss.

[leeperry]

Yeah I was recently wondering the same stuff as this was discussed ages ago and 16/9 1440*1080 stuff is really a GPU hog here, but I'm quite sure that madshi wouldn't force double scaling anyway.

All this said I've got the automatic PS gamut mapping profiling stuff working in PotP(and also 4:3 AR for 720*576 movies) and I'm always stunned by how good the AVS script looks, getting accurate colors really seems to increase the subjective CR.....that's what struck me when I first tried it on a "Mitsubishi HC3100" ages ago and it's just as apparent on my CRT...mix it w/ the killer 24p smoothness of mVR + Reclock and movies just look so true to life http://forum.slysoft.com/images/smilies/agreed.gif

There's only one last slight bug in PotP's automatic rules that would need to be ironed out but I guess I'm slowly running out of excuses to not ordering the BenQ W1070 at this point :devil:

I also see refurbished HD5870's being sold for 80€(w/ a 1y warranty), and that's quite tempting as it would appear to be almost as fast as the GTX660 (http://www.hardware.fr/articles/879-21/upgrade-face-x-anciennes-generations.html). I know that Jinc3 AR looks AMAZING, but tossing 200€ on a board that will be worth 30€ 2 years from now makes me feel conned :o

OTOH, I wanted to play around with 1080i@59.94fps and that'd be a no-go on the 5870's apparently..

[Joniii]

Using 1080p23 in display modes, changes to 24p even with PAL DVD's. Also using 720p50 or 1080p50 causes some occasional crashes when entering full screen or windowed.

[nevcairiel]

Surely 2x is rarely used for luma.

Like madshi pointed out, 2x will be useful in the future when watching 1080p on a 4K display (which happens to be 3840x2160, 2x 1080p)

So it may be rarely used today, but it may gain importance in the future.

[leeperry]

and here's a practical example of why 2X/3X TV/PC conversions were necessary, that's the original untouched DVD: http://www.discogs.com/Various-Burt-Sugarmans-The-Midnight-Special-1974/release/2274558

and now its ffdshow histogram: http://thumbnails103.imagebam.com/22929/c4256f229289037.jpg (http://www.imagebam.com/image/c4256f229289037)

results speak for themselves:

http://thumbnails108.imagebam.com/22929/208c12229288993.jpg (http://www.imagebam.com/image/208c12229288993) http://thumbnails104.imagebam.com/22929/f75000229289002.jpg (http://www.imagebam.com/image/f75000229289002) http://thumbnails108.imagebam.com/22929/e94ef6229289011.jpg (http://www.imagebam.com/image/e94ef6229289011)

http://thumbnails104.imagebam.com/22929/9db825229289017.jpg (http://www.imagebam.com/image/9db825229289017) http://thumbnails107.imagebam.com/22929/bbea0e229289025.jpg (http://www.imagebam.com/image/bbea0e229289025) http://thumbnails103.imagebam.com/22929/87a147229289031.jpg (http://www.imagebam.com/image/87a147229289031)

vintage tape encodes often end up sloppy like this IME, I guess that's what happens when careless engineers work on the case..maybe these were 0-255 in the first place and went through several PC>TV conversions duh :o

PQ is too good to be true in deinterlaced 59.94fps / Jinc3 AR / proper gamut mapping & the right levels conversion of course :cool:

[pie1394]

Yeah I was recently wondering the same stuff as this was discussed ages ago and 16/9 1440*1080 stuff is really a GPU hog here, but I'm quite sure that madshi wouldn't force double scaling anyway.


Many terrestrial-broadcasted HDTV programs are still encoded with 1440x1080 16:9 AR mode due to channel's bandwidth allocation consideration. For those very computation intensive algorithms, I guess it won't be a waste of time to design a dedicated optimization for such 4:3 H-scaling factor on Luma and 8:3-H + 2:1-V on Chroma to be displayed at 1920x1080.


With Jinc3+AR scaling + Per-pixel-adaptive deinterlacing, GTX260+ is not even powerful enough for any content displayed at 1920x1080. I have to set madVR to use Lanczos3+AR or turn off Anti-Ring option.

The newest madVR's Jinc3+AR 2x mode performance improvement now makes GTX260+ fast enough to handle those interlaced 1920x1080i30 4:2:0 contents displayed at 1920x1080p60. Without this improvement, the GTX260+ is even not fast enough to handle such content with only Chroma Jinc3+AR scaling.

Thus I wonder how many GPUs are really powerful enough to handle the critical case with Jinc3+AR mode?

With the progressive contents, GTX260+ is still quite capable of handling various materials with Jinc3+AR. There is no need to tell about newer GPUs with similar or more stream processor units.


The GT200b's (Tesla core) CUDA unit design is not as efficient as GF1xx's (Fermi core) or GK10x's (Kepler core). I just hope such old-school GPU is still capable of handling Jinc3+AR scaling on most cases w/o sacrificing nVidia's post-processing quality on HDTV contents.

[6233638]

I don’t know that I would spend €200 on a new video card just for Jinc AR scaling, especially if something like Lanczos 3 AR is an option for you.

I think most of us with higher-end video cards probably bought them for other purposes (most likely gaming) with madVR being a secondary concern.

Who knows if the card you go out and buy for Jinc 3 will be enough for some future developments, or if madVR will get a lot more efficient in the future? (E.g. Hard-coded optimisations for common sizes like 360/480/576/720 scaling, to 1080)

[DragonQ]

Like madshi pointed out, 2x will be useful in the future when watching 1080p on a 4K display (which happens to be 3840x2160, 2x 1080p)

So it may be rarely used today, but it may gain importance in the future.
True.

Tis still depressing that I can't use any of the higher quality algorithms on a current generation IGP. I guess Trinity would be better, as will Haswell.

[DragonQ]

Many terrestrial-broadcasted HDTV programs are still encoded with 1440x1080 16:9 AR mode due to channel's bandwidth allocation consideration.
In which countries? I know the UK doesn't use 1440x1080 any more and, AFAIK, the US never did since they seem to prefer 720p/60 when bandwidth is limited.

[aufkrawall]

I found each seperate device for presentation and DXVA processing helpful for 4k content and DXVA decoding: Without, render queue is very unstable.
No problems with software decoding, then it can be off.

Edit: I also think presented frames in advance should be set to 8 (at least) by default, 4 is definitely too few for 4k or other heavy stuff (10 bit I444 etc).

[Mangix]

http://www.monitortests.com/forum/Thread-Custom-Resolution-Utility-CRU?pid=105#pid105

It seems that even though madVR is telling D3D(or whatever) to switch to 60Hz, it ends up switching to 59Hz. Definite bug.

[DragonQ]

Pretty sure that's a Windows 8 bug.

[kerman]

It says to support with external dxva decoders, but if I enable coreavc dxva acceleration I get a full green screen... is it a bug??

[e-t172]

http://www.monitortests.com/forum/Thread-Custom-Resolution-Utility-CRU?pid=105#pid105

It seems that even though madVR is telling D3D(or whatever) to switch to 60Hz, it ends up switching to 59Hz. Definite bug.

"It's not a bug, it's a feature!" (http://support.microsoft.com/kb/2006076)

[turbojet]

Is there any way to play 2 videos simultaneously with dxva2 resizing without a bunch of dropped frames?
It's not a load issue, 2 videos with Jinc work fine on same hardware.

[pie1394]

In which countries? I know the UK doesn't use 1440x1080 any more and, AFAIK, the US never did since they seem to prefer 720p/60 when bandwidth is limited.

You are right. it is not so widely used now...

So what format does the BBC1 HD (DVB-T2 8MHz ?) use now?

I just realized that I have not watched any BBC One HD sample made from 2012. The latest one I seen was from 2011.
It looks like all BBC HD contents have been sent in 1920x1080 format since May 30, 2012...
The experiments were made since 2011 Wimbledon...


As I can tell, the Japan (ISDB-T 6MHz), Austrilla (DVB-T 7MHz?) terrestrial programs based on MPEG-2 video, they are still 1440x1080.

About the Korea (ATSC, DMB), 1920x1080 MPEG-2...

About the China (DMB), its HDTV program is 1920x1080 although still encoded with MPEG-2 format.

About the Hong Kong (DMB-T 8MHz) and Taiwan (DVB-T 6MHz), the HDTV programs are 1920x1080 encoded with H.264.

[pie1394]

First of all, thanks for your continued hard work on MadVR. A really good piece of software. :cool:

I've encountered a rather strange bug, happens randomly and with quite specific settings. Here's how to reproduce:

Enable both "use separate device for presentation" and "use separate device for DXVA processing".
Set "DXVA2" for image up-/downscaling.
Play a h264 10bit file in windowed mode, while using downscaling (or upscaling).
The video is "jumping", as if instead of next frame, the previous frame gets rendered.
What I mean is instead of frames 1 2 3 4, you get frames 1 2 1 4 or something.


Same file works fine with one of the "use separate device ..." options disabled or when using either other scaling algorithms or no scaling. Also other files work perfectly.

<... cut ... >

Intel HD 4000 + drivers v2875
Windows 7 (x64)
MadVR v0.85.7
MPC-HC Lite 1.6.6.6391 (32bit of course)
LAV Filters 0.54.1, but also happens if using MadVR's internal decoding

1. This issue also randomly happens to my ION system with the same setup, especially at 50% zoom window mode.
2. Driver version does not seem to matter. GeForce driver 306.97 or 310.70.
3. Not only H.264 Hi10p contents, it also happens to generic Bluray 24fps contents.
4. Same contents are repeatly verified with Bilinear luma scaling or MPC-BE64 + EVR. They just look ok.


The madVR's CPU usage is still somewhat higher than the EVR's. The difference is very obvious with the FHD H.264 Hi10p contents.

With P10 input contents, DXVA scaling causes higher CPU loading than Bilinear scaling on 1920x1080p24 --> 960x540p60 at window mode. This makes the C2D P8600@2.4GHz overloaded and it causes frame dropping. Change the Luma scaling to Bilinear, and it is about 75 ~ 90%. If MPC-BE64 + EVR is used, the CPU loading is just within 45 ~ 60% range.

Maybe the 64-bit Lav video decoder has some advantages over the 32-bit one due to compiler optimization from 8 additional GPRs in x86-64 mode.

[Mangix]

"It's not a bug, it's a feature!" (http://support.microsoft.com/kb/2006076)

which basically means that in my case, defining a resolution of 59.94Hz and not one of 60.00Hz is the only way to "fix" this. Oh well.

[annovif]

Like madshi pointed out, 2x will be useful in the future when watching 1080p on a 4K display (which happens to be 3840x2160, 2x 1080p)

So it may be rarely used today, but it may gain importance in the future.

Hello, a curiosity : Do you know why the marketing tell us that 4k is four time high definition ? 1080p*4 is a highest resolution than the sony 1000 4k matrix has .

Sorry for my very bad English
Good Year, Fabio

[nevcairiel]

Do you know why the marketing tell us that 4k is four time high definition ?

Because it is.
You can fit four 1080p images into one 4k image.

http://www.rtings.com/images/resolutions-ultra-hd-4k-1080p-720p-dvd.png

The sides are only 2x the size of course, but the overall image area is 4 times 1080p, which means the number of pixels increases by a factor of 4.

Do the math:
3840*2160 = 4*1920*1080

[jmone]

Consumers are so used to MegaPixel count I wish they would just stick to this as "4K" is really 8MP.

DVD = 0.3MP (NTSC) or 0.4MP (PAL)
720p = 0.9MP
1080p = 2MP
4K = 8MP

[shawkie]

I'm running MPC-HC + madVR on an HD4000. At the moment I have disabled both the internal MPEG2 decoders in MPC-HC and enabled the Intel decoders in madVR instead. I'm having a couple of problems with DVD menus. Firstly, I'm not getting any kind of highlight on the selected menu item - this is going to make it impossible to navigate the menu using a remote control. Secondly, on "The Business" (UK) the menu doesn't respond to the mouse either.

[DragonQ]

You are right. it is not so widely used now...

So what format does the BBC1 HD (DVB-T2 8MHz ?) use now?

I just realized that I have not watched any BBC One HD sample made from 2012. The latest one I seen was from 2011.
It looks like all BBC HD contents have been sent in 1920x1080 format since May 30, 2012...
The experiments were made since 2011 Wimbledon...
Correct, just before the Jubilee celebrations the final 1440x1080i channels (BBC HD & BBC One HD on satellite, plus BBC HD, BBC One HD, ITV1 HD & Channel 4 HD on terrestrial) permanently switched to 1920x1080i.

As I can tell, the Japan (ISDB-T 6MHz), Austrilla (DVB-T 7MHz?) terrestrial programs based on MPEG-2 video, they are still 1440x1080.

About the Korea (ATSC, DMB), 1920x1080 MPEG-2...

About the China (DMB), its HDTV program is 1920x1080 although still encoded with MPEG-2 format.

About the Hong Kong (DMB-T 8MHz) and Taiwan (DVB-T 6MHz), the HDTV programs are 1920x1080 encoded with H.264.
I imagine it's more likely to be seen in countries that still use MPEG2 for HD, except those that prefer 720p like the USA.

[shawkie]

I'm running MPC-HC + madVR on an HD4000. At the moment I have disabled both the internal MPEG2 decoders in MPC-HC and enabled the Intel decoders in madVR instead. I'm having a couple of problems with DVD menus. Firstly, I'm not getting any kind of highlight on the selected menu item - this is going to make it impossible to navigate the menu using a remote control. Secondly, on "The Business" (UK) the menu doesn't respond to the mouse either.

Correction: "The Business" (UK) does respond to the mouse (I just hadn't found the right place to click).

Also, I've noticed that if I install the LAV Filters then the highlights start getting displayed. This happens even though I haven't configured either MPC-HC or madVR to use them. Does anyone know what is going on?

[Luv]

Madshi !
Grateful thanks for your awesome software.I'm using 85.4 right now,with Spline 4 taps,HD 4770,Catalyst 12.4,XP32.And it's working perfectly.

PS:
Madshi,there's a Intel Media SDK 2012 available.Is there something that needs to be updated (I have the Intel MPEG2 decoder enabled)?

[Niyawa]

Hey people. For those who remember me, I made a graphic based on madVR algorithms to help make a range between the "Lowest" and "Highest" (previously "Ultra", heavy word). You can see it below.

http://i.imgur.com/8pNQv.png

Thanks for madshi and 6233638 help, I've made the one above with the most accuracy possible based on their respective quality x performance range. "Low" is also madVR default now.

Now there's something different, again, I'm not making those graphic out of nothing, but because we want to implement it in KCP (http://haruhichan.com/forum/showthread.php?7545-KCP-%93Kawaii-Codec-Pack%94) pack. Thus, now I'm looking for a graphic that can give people a decent reference on what their hardware can play, look below.

http://i.imgur.com/YCyJ7.png

It's far from accurate especially the GPU section. I want to improve it but it's based on research and I'm limited to my own hardware. That's why I want everyone to contribute a little by giving me some feedback in what they are or not able to play without problems. Those who have less powerful GPUs are the one I'm looking for. Far as I know, even a GTX 260 can play the "Highest" without problems, but even still, if there's a GPU even less powerful that can still play something at some range, I'd like to know. Anything is appreciated.

Oh yeah, thanks madshi for your response regarding buffers and flush settings, I missed that for some reason. I still want to know about the overlay bug thing, you said you would look up into it. Just to add more, it seems that it also changes my video brightness when it's enabled (making it somewhat more bright), and it seems it's not something normal. Not to mention it disable some effects in Windows 8 (like the little shadow effect in the windows). I tried to make some samples but I can't get them with overlay enabled. Remember that I'm using Intel GMA HD4500 (is overlay even compatible with it?)

If anyone can help (or teach me like the last time), I'll appreciate it as well.

[shimaflarex]

http://i.imgur.com/YCyJ7.png

CPU clock, VRAM, RAM, RAM speed have little (if any) influence on which preset should be chosen.

Things like GPU architecture, number of processing units on the GPU, GPU clock, GPU shader clock, GPU memory speed, GPU memory bus length etc are the parameters that matters the most, but how much each of those things affect the performance depends on the GPU manufacturer/architecture...

There's no way to create a table like the one you are trying to...The only viable approach is doing some sort of specific GPU benchmark.

Edit:Quoted the wrong table :P

[Niyawa]

CPU clock, VRAM, RAM, RAM speed have little (if any) influence on which preset should be chosen.
Not exactly, RAM indeed does not make too much difference when you look at it, but the clock does. That is noticeable when you have a laptop for example, maybe not too much on a Desktop though. I know someone who had a dual with 2.0 GHz but couldn't play 1080p with 10-bit properly. The only thing he did was upgrade to a core quad i5 (Ivy) and does that flawlessly now. He used to have 100% of use but now it's only around 20-35%. That is he using madVR.

Things like GPU architecture, number of processing units on the GPU, GPU clock, GPU shader clock, GPU memory speed, GPU memory bus length etc are the parameters that matters the most, but how much each of those things affect the performance depends on the GPU manufacturer/architecture...
That is what makes it more difficult. For example, I can make a range of what can or not play something, but it's hard to make good one. I know my Intel can play "Lowest" without problems, but "Low" ahead I have to tune a little. GTX 260 can play "Highest" without problems, so I guess it's a starting point in what to look up.

There's no way to create a chart like the one you are trying to...The only viable approach is doing some sort of specific GPU benchmark.
I don't want to make a perfect or "super" accurate graph, but something we can use for reference. Something someone can look at and say "Oh, I guess I can play that". I've already did benchmark regarding the "Middle" to "Lowest", so the only one I really want is to know what can play "Highest" and "High" at the lowest starting point.

Again, that is only possible not only with benchmarking but with old or simply not really powerful GPUs.

[shimaflarex]

Not exactly, RAM indeed does not make too much difference when you look at it, but the clock does. That is noticeable when you have a laptop for example, maybe not too much on a Desktop though. I know someone who had a dual with 2.0 GHz but couldn't play 1080p with 10-bit properly. The only thing he did was upgrade to a core quad i5 (Ivy) and does that flawlessly now. He used to have 100% of use but now it's only around 20-35%. That is he using madVR.

He couldn't play Hi10p 1080p content because decoding those is very CPU intensive. It wouldn't matter if he was using the "Lowest" or the "Highest" profile, changing the scaling algorithm doesn't affect CPU usage at all, as those are implemented using GPU shader code. And no, RAM clock barely affects anything at all, and even if it did, there is the issue of which is better: higher clock or lower latency. It can affect the CPU decoding efficiency, but again, it is irrelevant to choosing the preset.

fyi, I have a 2.4Ghz 4-core CPU, 4GB of RAM, with a 512MB video card, and it is unable to use your "Low" preset.

[ralle_h]

Hey madshi,

first of all I wanted to thank you for your great and continuous work on the best renderer out there. I really appreciate your efforts!

However, I'm having issues with madVR on one of my setups (on a lot of other desktop and HTPC's I did build there are no problems, so I doubt it is the configuration), which is an i3 2120 with the HD2000 IGPu, 8GB ram on a H77 PantherPoint Mainboard, running Windows Server 2008 R2. Furthermore I use the latest version of MPC-BE and LAV Splitter + Decoders for playback.

I think the OS could be the problem here, but I'm not sure and would appreciate any hint or tip coming from your side (or any other user who had similar issues or ideas about this).

1. Even though it's not true, the madVR menu shows that 2 instances of the player (in my case MPC-BE):

http://www.abload.de/img/madvr_bug8purw8vbjx.png

MPC-BE<$8dc>
MPC-BE<$8dc>

2. At times I'm getting random crashes when I open any files. If I open the file again, it works, but sometimes it doesn't - it's totally random.

Here is a part of the error log:

date/time : 2013-01-01, 23:15:04, 74ms
computer name : BOSS-SERVER
user name : Administrator <admin>
registered owner : Windows-Benutzer
operating system : Windows 2008 R2 x64 Service Pack 1 build 7601
system language : German
system up time : 40 seconds
program up time : 763 milliseconds
processors : 4x Intel(R) Core(TM) i3-2120 CPU @ 3.30GHz
physical memory : 6670/7883 MB (free/total)
free disk space : (C:) 61,37 GB
display mode : 1920x1080, 32 bit
process id : $e54
allocated memory : 356,52 MB
command line : "C:\Program Files (x86)\MPC-BE\mpc-be.exe" "D:\files\test.mkv"
executable : mpc-be.exe
current module : MADHCNET.DLL
module date/time : 2012-09-23 10:14
version : 1.0.11.0
compiled with : Delphi XE
madExcept version : 4.0.5
callstack crc : $73622432, $d398d782, $861e5b99
exception number : 1
exception class : Exception
exception message : Access violation at address $73622432 in module 'd3d9.dll'. Read of address $0.

CFrameQueue::RenderThread ($f90): <priority:2>
73622432 +0000 d3d9.dll
4a462469 +0049 madVR.ax rendering.cpp 150 +4 CRendering.CreatePixelShader
4a46b234 +3fe4 madVR.ax rendering.cpp 2477 +777 CRendering.Config
4a42d958 +2398 madVR.ax framequeue.cpp 8800 +520 CFrameQueue.RenderThread
4a40fdd6 +0006 madVR.ax framequeue.cpp 165 +0 Queue_RenderThread
75bc33a8 +0010 kernel32.dll BaseThreadInitThunk

main thread ($e58):
758a7908 +26 USER32.dll GetMessageW
75bc33a8 +10 kernel32.dll BaseThreadInitThunk

I would gladly email/post the whole log to you (via PM), if you want me to.

If you have any questions about my system please feel free to ask!

Kind regards,

[baii]

It highly dependent on the content type, maybe you should make separate tables.

CPU indeed does not matter if you do not use DXVA modes, but it is kind of tricky if you do `~
Also putting CPU into levels is even harder than GPU due to architecture changes ~~.

[Mangix]

Not sure if it's pushing it but I would bump up chroma to Jinc on high. I currently run Jinc3/Lanczos3/Catmull-Rom on a GTS 450. Jinc3 works for luma but gets frame drops.

[SamuriHL]

I run Jinc3 AR/Jinc3 AR/Lanczos4 AR LL on my 450. No frame drops. That's with DXVA CB in LAV filters and MC18 on an i5-2400s.

[jmone]

I run Jinc3 AR/Jinc3 AR/Lanczos4 AR LL on my 450. No frame drops. That's with DXVA CB in LAV filters and MC18 on an i5-2400s.

And then it get more complicated as this setup is fine... till you get 1080/60i material and with HW Deinterlacing it pushed it over the edge. I had to move up to the GTX660 to get all formats I have playing with Jinc3 AR (it was the AR that pushed it too hard on the 60i stuff).

[SamuriHL]

Except that I've done 1080/60i on that setup. :) No frame drops. Perfectly smooth. Obviously deinterlacing with madVR.

[dansrfe]

So, while I've been looking at GTX 680s and GTX 690s this weekend I decided to reformat my other desktop build that has an old GeForce 9500GT card and installed the latest builds of MPC-HC, LAV Filters, and madVR and set chroma upscaling to Bicubic 75 with AR and image upscaling to Jinc 3 without AR. Previously I was restricted to having AR disabled and using Lanczos 3 for image upscaling in order to not drop frames.

Three questions:

1) What is benefit in buying a high-end GTX 6xx card now vs buying a much cheaper GTX 4xx or 5xx series since those would certainly be able to play videos at max settings if my 9500 GT can.

2) Why do some users report having frame drops with a GTX 4xx or GTX 5xx with max settings?

3) What should I be looking for in a GPU with respect to madVR's processing/scaling/etc. needs? I've seen many benchmarks and read gaming forums in which some people say a variety of different things in a graphics card are equally important and others say that some things are more important than others when looking for a graphics card.

[baii]

Newer card have support for newer directx features for games, also they tend to be more power efficient.

4xx or 5xx doesn't have real meaning, it is the second letter of the card model that determine if it is a high end card. Also contents and setting have alot to do with frame drop. Everyone is playing different content in different setups.

[Niyawa]

He couldn't play Hi10p 1080p content because decoding those is very CPU intensive. It wouldn't matter if he was using the "Lowest" or the "Highest" profile, changing the scaling algorithm doesn't affect CPU usage at all, as those are implemented using GPU shader code. And no, RAM clock barely affects anything at all, and even if it did, there is the issue of which is better: higher clock or lower latency. It can affect the CPU decoding efficiency, but again, it is irrelevant to choosing the preset.

fyi, I have a 2.4Ghz 4-core CPU, 4GB of RAM, with a 512MB video card, and it is unable to use your "Low" preset.
That makes much more sense, but then, is the CPU table even usable at all? I should remove the RAM I guess. Should we just rely on the GPU then?

It highly dependent on the content type, maybe you should make separate tables.

CPU indeed does not matter if you do not use DXVA modes, but it is kind of tricky if you do `~
Also putting CPU into levels is even harder than GPU due to architecture changes ~~.
I know it really depends a lot. There's some hardware out there that even though is 4-6 years old can still playback "Highest" at it's finest, but mine that is only 2-3 can't even do Middle decently. Of course, I'm aware of different architectures.

Not sure if it's pushing it but I would bump up chroma to Jinc on high. I currently run Jinc3/Lanczos3/Catmull-Rom on a GTS 450. Jinc3 works for luma but gets frame drops.
My memory fails a little, but I remember someone (I guess it was madshi) saying that Jinc is too GPU intensive compared to the others, so putting it on "High" would be asking for too much. Again, it depends a lot on the hardware, but going for the actual reference, it's better to have that on the "Highest" only.

I run Jinc3 AR/Jinc3 AR/Lanczos4 AR LL on my 450. No frame drops. That's with DXVA CB in LAV filters and MC18 on an i5-2400s.
I want to believe you. But...

And then it get more complicated as this setup is fine... till you get 1080/60i material and with HW Deinterlacing it pushed it over the edge. I had to move up to the GTX660 to get all formats I have playing with Jinc3 AR (it was the AR that pushed it too hard on the 60i stuff).
This happens. And-

Except that I've done 1080/60i on that setup. :) No frame drops. Perfectly smooth. Obviously deinterlacing with madVR.
We get to this again.

When people said to me that making a graph would be difficult, I believed in them, but I'm starting to think even a reference one would be impossible too.

1) What is benefit in buying a high-end GTX 6xx card now vs buying a much cheaper GTX 4xx or 5xx series since those would certainly be able to play videos at max settings if my 9500 GT can.
This is just my pick, but the improved performance and less energy power needed is a good combination.

2) Why do some users report having frame drops with a GTX 4xx or GTX 5xx with max settings?
We can't really answer that, it depends a lot on the user's hardware.

3) What should I be looking for in a GPU with respect to madVR's processing/scaling/etc. needs? I've seen many benchmarks and read gaming forums in which some people say a variety of different things in a graphics card are equally important and others say that some things are more important than others when looking for a graphics card.
It also depends, most of graphic cards out there are focused on game performance. So we use benchmarks with fps to see which one gives better results... if we were to use a benchmark for madVR instead, I guess the rendering time (ms) would be the main factor? There's seems to be a limit here where I can use while not having dropped frames, but that's just my assumption based on my own eyes, not really numbers... which is bad in this situation.

[cyberbeing]

@Niyawa

Minimum VRAM should be based on the user's desktop resolution, and will always be the same for all the presets except lowest (Bilinear). If you want a basic guideline, it should be something like the following for optimal performance:

1366x768: 256MB VRAM
1600x900: 384MB VRAM
1920x1080: 512MB VRAM
2560x1440: 786MB VRAM

As for VRAM bandwidth, 15GB/s is around the bare minimum needed for playback, but you probably want 20GB/s+ to give a bit of performance headroom.

That said, if you want to automatically pick madVR settings presets for a user, it can't be done with a simple chart like you have. You either will to build a GPU database, and/or calculate theoretical Pixel Fillrate, Texel Fillrate, and Memory Bandwidth values on the fly. Shader performance determines 99% of madVR performance, yet you've left it out of your chart entirely. Alternatively, you could wait for madshi to release an updated madVR benchmark, and create a DB that way. Also keep in mind that pushing the limits of what a GPU is capable of with 24fps video is a bad idea, since it will result in completely unwatchable 60fps with constant dropped frames. For example, there are still a minority of anime Blu-rays which are released as 1080i, which result in 24fps / 60fps VFR.

Without doing the above, a codec pack should only be deciding if a user needs to use Bilinear-only, instead of the madVR defaults, to help ensure reliable playback on install. For other more intensive & high-quality madVR presets, it would be best to have the user manually toggle them in the codec pack's settings app after install. At that point, if they don't work well, it would be the user's fault.

[pie1394]

Except that I've done 1080/60i on that setup. :) No frame drops. Perfectly smooth. Obviously deinterlacing with madVR.

1440x1080i60 --> 1920x1080p60 ?

If it still does not cause any frame dropping with Jinc3+AR mode, it means the Fermi core's efficiency is much better than the Tesla core's. Or the different per-pixel-adaptive deinterlacing algorithms are performed on these two GPUs.

Your GTS450 : 783 MHz 192 core + 128-bit DDR5-3600

My GTX260+ : 680 MHz 216 core + 384-bit DDR3-2000

[ryrynz]

Madshi, can overlay's transition to windowed mode fullscreen be improved? The windowed image tends to stick on screen for a bit, with the enlarged image sometimes shifting back and forward a few frames before playing normally whilst the windowed image is still visable in the top left corner. FSE mode when enabled provides a much smoother transition without any issues. On HD3000/4000 7 x64.

[Niyawa]

Minimum VRAM should be based on the user's desktop resolution, and will always be the same for all the presets except lowest (Bilinear). If you want a basic guideline, it should be something like the following for optimal performance:

1366x768: 256MB VRAM
1600x900: 384MB VRAM
1920x1080: 512MB VRAM
2560x1440: 786MB VRAM
Thanks, I'll be using that info.

As for VRAM bandwidth, 15GB/s is around the bare minimum needed for playback, but you probably want 20GB/s+ to give a bit of performance headroom.
Thanks for your thoughts on this as well.

That said, if you want to automatically pick madVR settings presets for a user, it can't be done with a simple chart like you have. You either will to build a GPU database, and/or calculate theoretical Pixel Fillrate, Texel Fillrate, and Memory Bandwidth values on the fly. Shader performance determines 99% of madVR performance, yet you've left it out of your chart entirely. Alternatively, you could wait for madshi to release an updated madVR benchmark, and create a DB that way. Also keep in mind that pushing the limits of what a GPU is capable of with 24fps video is a bad idea, since it will result in completely unwatchable 60fps with constant dropped frames. For example, there are still a minority of anime Blu-rays which are released as 1080i, which result in 24fps / 60fps VFR.
Actually, I didn't "left-out". I was going to do more research in GPU clock and other matters including this one, you gave me the heads up early then I thought, haha. As for the 24fps, I have some samples with 60 fps here, but I admit I never got around testing them. Now a database huh.... well. Doing some filtering and some more research I could make a list of supported GPUs and their respective presets, the problem will be how we'll get that working and what algorithm we will use, since we still don't know what's the GPU at lowest range that can play "Highest" preset. We could work around that later thought.

Without doing the above, a codec pack should only be deciding if a user needs to use Bilinear-only, instead of the madVR defaults, to help ensure reliable playback on install. For other more intensive & high-quality madVR presets, it would be best to have the user manually toggle them in the codec pack's settings app after install. At that point, if they don't work well, it would be the user's fault.
That's a pretty good point, but we did those presets (the algorithm one) to make a convenient yet good range of quality x performance in one-click. For now that's what we've been doing: make the user select what he wants, "Middle" is default since most modern machines can play it (though I'm starting to doubt it). We're already working around a tool that resets the presets without needing to re-install, so soon enough it won't be a problem if he selects something that it's not up to the job.

Hell, I have to agree with basically most of what you said. For now I should put the graph in On-Hold then. We'll get around that when madVR is close to 1.x and the pack is more user-friendly. But thanks cyber and everyone, you guys have been a big help. Don't mind if I come here again with some more things to discuss, you all are everything I got.