madVR - high quality video renderer (GPU assisted)

[leeperry]

Originally Posted by madshi

@leeperry? (Question is about how to contact PotPlayer developers.)

like this, but I'm 99.99% sure that the answer will be "use other player!!!!!!!!!!!"

so he'd better make a compelling case

[HoP]

Originally Posted by madshi

EVR exposes different interfaces to the media player compared to VMR9. Try VMR9, I think it will behave similar to madVR. madVR currently exposes the same interfaces that VMR9 exposes. I plan to add support for EVR interfaces in a future version, but probably not too soon.

thanks for info
yes,"VMR9 windowed" is similar to madVR.
but "VMR9 renderless" is not.

[madshi]

madVR v0.85.5 released

http://madshi.net/madVR.zip

* fixed: DXVA deinterlacing: when GPU couldn't keep up, audio sync got lost
* fixed: odd source rectangles could result in "green" images (DXVA)
* fixed: display mode changer sometimes didn't work for DVDs
* slightly modified "use separate device for DXVA processing" behaviour
* changed scaling defaults once more (Bicubic75, Lanczos3, Catmull-Rom)
* disabled "perform deinterlacing in separate thread" by default
* DXVA NV12 conversion routine now uses 16bit float instead of 32bit (faster)
* up to 56% speed improvement for Jinc3    chroma upscaling
* up to 40% speed improvement for Jinc3 AR chroma upscaling
* up to 53% speed impr. for Jinc3/4     image upscaling with 2x scaling factor
* up to 47% speed impr. for Jinc3/4 AR  image upscaling with 2x scaling factor
* up to 39% speed impr. for non-Jinc    image upscaling with 2x scaling factor
* up to 27% speed impr. for non-Jinc AR image upscaling with 2x scaling factor
* up to 44% speed impr. for non-Jinc    image upscaling with 3x scaling factor

Please note that the speed improvements only apply in very specific circumstances. E.g. the Jinc chroma upsampling speed improvement only applies to 3-taps (with or without AR). There's no improvement for Jinc4/8 chroma upscaling. And all the other improvements only apply if there's an *exact* 2x or 3x scaling factor requested by the media player. With any other scaling factor there's no improvement at all for image scaling performance. The exact scaling factors of 2x and 3x allow me to write hard wired shader code which runs more efficiently, thus the speed improvement. There are no further similar improvements for other scaling factors planned. The 3x scaling factor improvements only apply when not using AR. There's no improvement for 3x AR image upscaling.

The main purpose of the 2x and 3x speed improvements is the upcoming UltraHD (4k) displays. Now with v0.85.5 you can upscale Blu-Ray to UltraHD with significantly lower GPU usage, as long as you use exactly 2x upscaling. Also 720p -> UltraHD happens to be exactly 3x upscaling, so that's why I added some dedicated 3x pixel shaders, too. Please note that the exact speed improvement depends on the circumstances and especially on the GPU model.

Here's the list of my own benchmarks, tested with NVidia 9400 (very old and slow) and AMD 7770:

2x scaling
tap 2 :  8.53 ->  6.51  (23.7%) NVidia 9400
tap 3 : 11.98 ->  7.86  (34.4%) NVidia 9400
tap 4 : 14.92 ->  9.02  (39.5%) NVidia 9400
tap 2 :  6.30 ->  5.92  ( 5.9%) AMD 7770
tap 3 :  8.08 ->  6.02  (25.5%) AMD 7770
tap 4 :  9.55 ->  6.75  (29.3%) AMD 7770

3x scaling
tap 2 : 13.16 ->  9.91  (24.7%) NVidia 9400
tap 3 : 19.96 -> 12.65  (36.6%) NVidia 9400
tap 4 : 25.24 -> 17.08  (32.3%) NVidia 9400
tap 2 :  9.04 ->  7.61  (15.8%) AMD 7770
tap 3 : 12.12 ->  7.84  (35.3%) AMD 7770
tap 4 : 16.58 ->  9.28  (44.0%) AMD 7770

2x AR scaling
tap 2 : 16.75 -> 13.64  (18.6%) NVidia 9400
tap 3 : 20.24 -> 17.79  (12.1%) NVidia 9400
tap 4 : 24.78 -> 21.87  (11.7%) NVidia 9400
tap 2 :  7.99 ->  6.45  (19.3%) AMD 7770
tap 3 :  9.71 ->  7.29  (24.9%) AMD 7770
tap 4 : 11.47 ->  8.36  (27.1%) AMD 7770

2x Jinc scaling
tap 3 : 38.84 -> 19.23  (50.5%) NVidia 9400
tap 4 : 61.27 -> 40.74  (33.5%) NVidia 9400
tap 3 : 14.09 ->  7.33  (48.0%) AMD 7770
tap 4 : 25.55 -> 12.00  (53.0%) AMD 7770

2x Jinc AR scaling
tap 3 :  86.92 ->  70.58  (18.8%) NVidia 9400
tap 4 : 137.87 -> 118.62  (14.0%) NVidia 9400
tap 3 :  15.97 ->   9.78  (38.8%) AMD 7770
tap 4 :  27.00 ->  14.30  (47.0%) AMD 7770

Jinc Chroma scaling
tap 3 : 38.84 -> 27.57  (29.0%) NVidia 9400
tap 3 : 14.27 ->  6.24  (56.3%) AMD 7770

Jinc Chroma AR scaling
tap 3 : 89.03 -> 86.84  ( 2.5%) NVidia 9400
tap 3 : 15.90 ->  9.50  (40.3%) AMD 7770

(I've used different test videos for AMD and NVidia GPUs, and also for different test sequences. So the absolute rendering times are not suitable for any sort of comparisons).

-------

Once again I need your FEEDBACK:

I'm trying to figure out again which defaults to use for a couple of options. I'm especially interested in these:

(1) "perform deinterlacing in separate thread". This was enabled in earlier madVR builds by default. It's now disabled by default in v0.85.5. With my AMD GPU, there's no difference. My NVidia GPU runs slightly better with this option disabled.

(2) "use a separate device for presentation". This was and still is enabled by default. It used to be a good method to reduce glitches with older NVidia drivers. I'm wondering if this option is still beneficial today with both NVidia and AMD GPUs. I've been told this option can make problems for NVidia Optimus, so I'm wondering whether I should disable it by default.

(3) "use a separate device for DXVA processing". This is a rather new option, introduced recently. I've got one report that it helped. So I wonder if I should enable it by default. But I guess this might also make problems with NVidia Optimus (because it's similar to (2), see above). Right now I'm wondering whether maybe I should combine options (2) and (3) into one, so that I'm either using only 1 device for everything, or 3 devices, one for rendering, one for presentation and one for DXVA processing. Thoughts?

(4) "enable windowed overlay". When I originally implemented overlay, it wasn't ready for prime time yet, so I disabled this by default. But since its introduction, I have applied lots of fixes, so now it seems to work very well (only for NVidia and Intel, unfortunately). It works so well IMHO that I'm reevaluating whether I should enable this by default (not for AMD, obviously). Any thoughts on this?

Please check which way these options work better for you, and let me know your test results. If you can't see a difference, please let me know that, too.

Thanks!

[madshi]

Originally Posted by HoP

yes,"VMR9 windowed" is similar to madVR.
but "VMR9 renderless" is not.

Yeah, madVR currently behaves like VMR9 windowed. In the future this might change, but that's the current state, and it's "as intended" for now.

[DrNein]

Originally Posted by madshi

(4) "enable windowed overlay". When I originally implemented overlay, it wasn't ready for prime time yet, so I disabled this by default. But since its introduction, I have applied lots of fixes, so now it seems to work very well (only for NVidia and Intel, unfortunately). It works so well IMHO that I'm reevaluating whether I should enable this by default (not for AMD, obviously). Any thoughts on this?

Thanks!

Another negative is that limits video to one at a time. Obscure perhaps but I frequently use multiple instances of MPC-HC. If enabled by default it would not really be obvious why that functionality stopped working.

No, thank you!

[madshi]

Originally Posted by DrNein

Another negative is that limits video to one at a time.

True, but I should be able to detect the situation and use standard windowed mode playback for the 2nd, 3rd etc instances of the media player.

[aufkrawall]

I can confirm that with 0.85.5 it now properly switches to the correct display mode with my tested DVD.

Furthermore, now it also seems to be fast enough for the 4k test sample @720p with native DXVA2, Jinc3 AR and CR AR LL, at least if I max out CPU, GPU and presentation queues (which doesn't seem to give me any real drawback).
At the moment it seems there is no single problem for me with this version on my system.
Boring.

[Majin3]

Originally Posted by madshi

(2) "use a separate device for presentation". This was and still is enabled by default. It used to be a good method to reduce glitches with older NVidia drivers. I'm wondering if this option is still beneficial today with both NVidia and AMD GPUs. I've been told this option can make problems for NVidia Optimus, so I'm wondering whether I should disable it by default.

(3) "use a separate device for DXVA processing". This is a rather new option, introduced recently. I've got one report that it helped. So I wonder if I should enable it by default. But I guess this might also make problems with NVidia Optimus (because it's similar to (2), see above). Right now I'm wondering whether maybe I should combine options (2) and (3) into one, so that I'm either using only 1 device for everything, or 3 devices, one for rendering, one for presentation and one for DXVA processing. Thoughts?

I get a black screen with any of them enabled with AMD's Switchable Graphics as well so it's not limited to Optimus.
I guess combining them sounds reasonable unless one of them has side effects (can't really notice any difference on my Nvidia desktop).

[Prinz]

With 0.85.5 I can now use DXVA native without framedrops without scaling. But only in windowed mode, in exclusive mode the render and present queue drop to 1 and 0. And the framedrops start... In windowed mode the all queues are full.

[Mangix]

Originally Posted by madshi

(4) "enable windowed overlay". When I originally implemented overlay, it wasn't ready for prime time yet, so I disabled this by default. But since its introduction, I have applied lots of fixes, so now it seems to work very well (only for NVidia and Intel, unfortunately). It works so well IMHO that I'm reevaluating whether I should enable this by default (not for AMD, obviously). Any thoughts on this?

I've used the Overlay mode since its inception on my GTS 450 and have never ran into any issues. The only issue with it is that when "present several frames in advance" is disabled, overlay is also disabled.

I vote for it to be enabled by default.

edit: I just noticed an issue when I disabled "use a separate device for presentation". With it enabled, my rendering times are around 8ms whereas with it disabled they are around 13ms. I am using CUVID as well as SVP. One of those could account for the difference. Not sure.

edit2: now this is odd. when I disable SVP, my rendering times go up to 17ms whereas with it enabled, i get 13ms... shouldn't a higher frame rate cause higher rendering times? I also disabled CUVID and went to software decoding and got 16ms instead of 17ms. So it's not an issue with CUVID.

[Niyawa]

@Kado and madshi
Thanks for the answers. Well, I wouldn't ask for something like that to the developers, and yes it works just fine right now. It was just a detail I wanted to know more about.

About the version 0.85.5, I tested this "Enable windowed overlay" and when I play a 8-bit video with DXVA2 enabled, (sometimes) the video start to shake (not all of it, usually half or 70-80% of the screen playing it). I have Intel HD 4500 and I'm using Windows 8 x64. I don't believe this is normal. If someone could explain to me what "exactly" this option does too, I would appreciate it, first time using it. Also, madshi, could you think of making a description for every option in madVR in the readme? If it's not asking too much of course, but there's a lot of things included in there that usually at least half most of beginners don't understand the meaning and it's point. For example, in this "faq" you would put a brief explanation in 1-2 lines maybe about what is the scaling algorithms, backbuffers and flush, how they work or how they affect playback.

Edit: Seems like this only happens with MPC-HC. I tried with PotPlayer but couldn't reproduce the problem there. I'm using MPC-HC Lite 1.6.6.6373.

[crotecun]

Originally Posted by aufkrawall

Yep, that madVR builds works and it works also with EVR half floating point processing (maybe a little worse).
With madVR I don't get any dropped frames after lowering queues in windowed mode.
But unfortunately, in FSE I get dropped frames, present queue sometimes gets emtpy.

Btw: I figured out why FSE wasn't working at all: It seems it didn't like 16 frames presented in advance, with lower values like 4 or 8 this doesn't happen. Then there are just those frame drops.

Originally Posted by madshi

Ah, that's interesting. I'm glad to hear that madVR seems to run just as fast as EVR with your ultra-slow GPU, when switching EVR to half-way decent quality settings. That does confirm that madVR can work with very low-power GPU hardware, too.

Which settings did he change to get it running nicely on an AMD E-350? I have given up on that graphics chip for madvr, I'm curious what could be done with the new build.

[6233638]

Originally Posted by madshi

The main purpose of the 2x and 3x speed improvements is the upcoming UltraHD (4k) displays. Now with v0.85.5 you can upscale Blu-Ray to UltraHD with significantly lower GPU usage, as long as you use exactly 2x upscaling. Also 720p -> UltraHD happens to be exactly 3x upscaling, so that's why I added some dedicated 3x pixel shaders, too.

Sounds great!

Originally Posted by madshi

(1) "perform deinterlacing in separate thread". This was enabled in earlier madVR builds by default. It's now disabled by default in v0.85.5. With my AMD GPU, there's no difference. My NVidia GPU runs slightly better with this option disabled.

I don't really see a difference with this on or off when using DXVA2 Copy-Back in LAV Video.

If I use DXVA2 Native though, there's a big difference:

• Enabled
• Disabled

Notice the queues, dropped frames, and render times.

Originally Posted by madshi

(2) "use a separate device for presentation". This was and still is enabled by default. It used to be a good method to reduce glitches with older NVidia drivers. I'm wondering if this option is still beneficial today with both NVidia and AMD GPUs. I've been told this option can make problems for NVidia Optimus, so I'm wondering whether I should disable it by default.

I'm getting much higher rendering times with this disabled, jumping from 24ms to 32ms average render times in Windowed mode with the video I tried, and 23ms to 29ms in Windowed Overlay.
In Fullscreen Exclusive mode, they jumped from 17ms to 27ms.

Originally Posted by madshi

(3) "use a separate device for DXVA processing". This is a rather new option, introduced recently. I've got one report that it helped. So I wonder if I should enable it by default. But I guess this might also make problems with NVidia Optimus (because it's similar to (2), see above). Right now I'm wondering whether maybe I should combine options (2) and (3) into one, so that I'm either using only 1 device for everything, or 3 devices, one for rendering, one for presentation and one for DXVA processing. Thoughts?

Using DXVA2 Native, this kills performance on my system using the GT610, whether the "use a separate device for presentation" option is enabled or not. (I also tried changing the "perform deinterlacing in a separate thread" option) 1080i50 is completely unplayable.

Originally Posted by madshi

(4) "enable windowed overlay". When I originally implemented overlay, it wasn't ready for prime time yet, so I disabled this by default. But since its introduction, I have applied lots of fixes, so now it seems to work very well (only for NVidia and Intel, unfortunately). It works so well IMHO that I'm reevaluating whether I should enable this by default (not for AMD, obviously). Any thoughts on this?

I really dislike not being able to grab a scaled screenshot with PrtScn when overlay is enabled. It's also a lot slower switching to fullscreen exclusive mode from windowed overlay. It's almost instantaneous switching from windowed mode if there's no refresh rate change (just flashes black for a fraction of a second) and the screen blacks out for much longer (a couple of seconds) going from windowed overlay to fullscreen exclusive, as if I had changed refresh rate.

I currently have CPU queue at 32 and GPU queue at 24 though, which may be slowing things down. (backbuffers/present queue at the default though)
Performance is slightly better, at 23ms rather than 24ms in windowed mode with the same sample, but doesn't seem worth it.

[madshi]

Originally Posted by aufkrawall

At the moment it seems there is no single problem for me with this version on my system.
Boring.

That sounds like the perfect time to do those tests for the feedback I need...

Originally Posted by Majin3

I get a black screen with any of them enabled with AMD's Switchable Graphics as well so it's not limited to Optimus.

Good to know, thanks.

Originally Posted by Prinz

With 0.85.5 I can now use DXVA native without framedrops without scaling. But only in windowed mode, in exclusive mode the render and present queue drop to 1 and 0. And the framedrops start... In windowed mode the all queues are full.

Interesting. Which GPU/OS? What happens if you disable/enable the "use a separate device for DXVA processing" and "perform deinterlacing in a separate thread"? Does that fix the problem with exclusive mode?

Originally Posted by Mangix

edit: I just noticed an issue when I disabled "use a separate device for presentation". With it enabled, my rendering times are around 8ms whereas with it disabled they are around 13ms. I am using CUVID as well as SVP. One of those could account for the difference. Not sure.

edit2: now this is odd. when I disable SVP, my rendering times go up to 17ms whereas with it enabled, i get 13ms... shouldn't a higher frame rate cause higher rendering times? I also disabled CUVID and went to software decoding and got 16ms instead of 17ms. So it's not an issue with CUVID.

The rendering times are per frame, and should be more or less independent of frame rate. Not sure exactly why your rendering times change, depends on those configuration changes you made. The bigger question would be how it affects real-life performance. E.g. you could up your scaling settings (e.g. switch to higher taps or to Jinc) to get your PC near to the max it can do. And then you could check whether changing those madVR option(s) helps or hurts.

Originally Posted by Niyawa

About the version 0.85.5, I tested this "Enable windowed overlay" and when I play a 8-bit video with DXVA2 enabled, (sometimes) the video start to shake (not all of it, usually half or 70-80% of the screen playing it). I have Intel HD 4500 and I'm using Windows 8 x64. I don't believe this is normal. If someone could explain to me what "exactly" this option does too, I would appreciate it, first time using it.

Edit: Seems like this only happens with MPC-HC. I tried with PotPlayer but couldn't reproduce the problem there. I'm using MPC-HC Lite 1.6.6.6373.

Hmmmm... I'm not sure what the media player has to do with this. It shouldn't really affect overlay playback. Sounds quite weird. Are you sure that madVR is really used by both media players? (You can turn on the debug OSD with Ctrl+J to make sure).

What you're describing as "shaking video": Do you mean tearing? Tearing means that one part of the image is one frame ahead of the rest of the image all the time. Or do you mean that frames appear to be shown out of order, so that movement goes back and forth all the time?

In any case, the problem goes away if you disable the Overlay option? Are you using DXVA decoding? Maybe it helps disabling that?

Originally Posted by Niyawa

Also, madshi, could you think of making a description for every option in madVR in the readme? If it's not asking too much of course, but there's a lot of things included in there that usually at least half most of beginners don't understand the meaning and it's point. For example, in this "faq" you would put a brief explanation in 1-2 lines maybe about what is the scaling algorithms, backbuffers and flush, how they work or how they affect playback.

I do plan to write a full documentation, but not right now. madVR is still not feature complete. Documentation will be done as the last step before releasing v1.0.

Originally Posted by crotecun

Which settings did he change to get it running nicely on an AMD E-350? I have given up on that graphics chip for madvr, I'm curious what could be done with the new build.

IIRC the main thing to do is to set all scaling options to Bilinear. Then you may want to turn on "use half framerate deinterlacing" and "run custom shaders in 16bit" in the "trade quality for performance" section. I guess that's it.

[madshi]

Originally Posted by 6233638

I don't really see a difference with this on or off when using DXVA2 Copy-Back in LAV Video.

If I use DXVA2 Native though, there's a big difference:

• Enabled
• Disabled

Notice the queues, dropped frames, and render times.

Ok, that's a really big difference. I really wonder why this option helps with your NVidia GPU, but hurts with mine. I haven't tested with DXVA decoding, though. I guess I should do that...

Originally Posted by 6233638

I'm getting much higher rendering times with this disabled, jumping from 24ms to 32ms average render times in Windowed mode with the video I tried, and 23ms to 29ms in Windowed Overlay.
In Fullscreen Exclusive mode, they jumped from 17ms to 27ms.

Ok. But do those rendering times also reflect real world? Meaning: Can you use higher scaling settings with "use a separate device for presentation" enabled?

Originally Posted by 6233638

Using DXVA2 Native, this kills performance on my system using the GT610, whether the "use a separate device for presentation" option is enabled or not. (I also tried changing the "perform deinterlacing in a separate thread" option) 1080i50 is completely unplayable.

Does this option also hurt when not using DXVA2 Native?

Originally Posted by 6233638

I really dislike not being able to grab a scaled screenshot with PrtScn when overlay is enabled. It's also a lot slower switching to fullscreen exclusive mode from windowed overlay. It's almost instantaneous switching from windowed mode if there's no refresh rate change (just flashes black for a fraction of a second) and the screen blacks out for much longer (a couple of seconds) going from windowed overlay to fullscreen exclusive, as if I had changed refresh rate.

I currently have CPU queue at 32 and GPU queue at 24 though, which may be slowing things down. (backbuffers/present queue at the default though)
Performance is slightly better, at 23ms rather than 24ms in windowed mode with the same sample, but doesn't seem worth it.

Interesting. For my old NVidia 9400, performance is *much* better in Overlay mode. Haven't checked the rendering times, but real life performance is so much better...

[ryrynz]

Madshi, Could we get dedicated code for 1.5x? 720 to 1080 is fairly common and this may help out some with fairly lowish end integrated graphics to use Jinc.
With these differences in scaling performances someone with border line graphics power may run into dropped frames when using the non hard wired code however.
That may be a good reason to allow for configurable resizers based upon tagging and input/output resolution.

[6233638]

Originally Posted by madshi

Ok. But do those rendering times also reflect real world? Meaning: Can you use higher scaling settings with "use a separate device for presentation" enabled?

Well, it means that more demanding videos are dropping frames, so I would have to lower my settings if I disabled "use a separate device for presentation".

Originally Posted by madshi

Does this option also hurt when not using DXVA2 Native?

Essentially no difference on or off when using DXVA2 Copy-Back.

Originally Posted by madshi

Interesting. For my old NVidia 9400, performance is *much* better in Overlay mode. Haven't checked the rendering times, but real life performance is so much better...

Even using Jinc 3 at a reasonably large scale factor (360p to 1080p) I'm only seeing a difference of about 1ms.

I used to see a lot of dropped frames in windowed mode for no apparent reason (low enough render times etc) but that's been fixed for a while now, so on my system there doesn't seem to be any real benefit to windowed overlay, and it brings some annoyances.

[madshi]

Originally Posted by ryrynz

Madshi, Could we get dedicated code for 1.5x? 720 to 1080 is fairly common and this may help out some with fairly lowish end integrated graphics to use Jinc.

I don't think I can do a faster hardwired Jinc shader for 1.5x. I've also not been able to do a faster 3x Jinc shader. Only 2x seems possible for Jinc, due to the way Jinc resampling works. Maybe it would be possible to create 1.5x hardwired shaders for non-Jinc scaling algorithms, but I'm not sure it's worth the effort...

Originally Posted by 6233638

Well, it means that more demanding videos are dropping frames, so I would have to lower my settings if I disabled "use a separate device for presentation".

Even using Jinc 3 at a reasonably large scale factor (360p to 1080p) I'm only seeing a difference of about 1ms.

You're only looking at the rendering times, though. I'm not convinced that they're *that* reliable that just comparing those numbers will tell you everything. They are somewhat reliable when changing scaling settings with everything else being the same. But if you change other options which modify the whole way madVR works, just looking at the rendering times alone might not paint the whole picture.

Originally Posted by 6233638

I used to see a lot of dropped frames in windowed mode for no apparent reason (low enough render times etc) but that's been fixed for a while now, so on my system there doesn't seem to be any real benefit to windowed overlay, and it brings some annoyances.

It's interesting that windowed mode playback seems to work better for you now. Do you have Aero on or off?

[6233638]

Originally Posted by madshi

You're only looking at the rendering times, though. I'm not convinced that they're *that* reliable that just comparing those numbers will tell you everything. They are somewhat reliable when changing scaling settings with everything else being the same. But if you change other options which modify the whole way madVR works, just looking at the rendering times alone might not paint the whole picture.

I agree, but when it's not dropping frames or showing other issues, I don't really have anything else to go on.

Originally Posted by madshi

It's interesting that windowed mode playback seems to work better for you now. Do you have Aero on or off?

I have Aero on. It's been fine for some time now though, I don't know if it was a driver change, or something in madVR that changed. I mostly watch in Fullscreen Exclusive mode though.

[cca]

Maybe it's me, but is the half-frame rate deinterlacing on by default now? How can I restore it to full frame ?

EDIT: No it wasn't it, I needed to activate deinterlacing in a separate thread for proper results.