madVR - high quality video renderer (GPU assisted)

[James Freeman]

Originally Posted by huhn

can you plz try this with 86.11 my version is crashing now and render times are about 10 ten higher than normal

Rendering times are exactly the same or a little longer with 86.11 (87.1b is faster) with 1080p24.

But changing to film mode with refresh change (86.11) crashes mpc-hc.exe, But 87.1b only blackens the image which is better.
Madshi, Its an OLD bug.

Overall, 87.1b is more stable and faster with my system.

[madshi]

Originally Posted by omarank

I just tried the latest test build of madVR (madVR871b). As others have reported I too see a regression in performace while playing interlaced videos as compared to v0.86.11.

So the 871b build did not help at all compared to the official v0.87.1?

Originally Posted by omarank

I have encountered one minor bug in the new release which I guess hasn't been reported. When I press the shortcut keys for <chroma upscaling algorithm - set to "Jinc">, it uses NNEDI3 algorithm and I see a green screen.

Thanks, confirmed.

Originally Posted by ryrynz

HD3000 on 8.1 with 3347 drivers using 0.87.1, when fullscreen exclusive is enabled and not ticking 'use separate device for presentation' and after transitioning to next file after playback I get a black screen.

Try the 871a or 871b test build.

[ryrynz]

Tried the a build. Fixed.

[James Freeman]

I just tried the latest test build of madVR (madVR871b). As others have reported I too see a regression in performace while playing interlaced videos as compared to v0.86.11.

I disagree.
87.1b works much faster than even 86.11.

[romulous]

Originally Posted by madshi

Maybe I could say something if I knew which GPU/OS you were using? Anyway, my priority right now is on making the "old" features work alright again. Only after that I'm ready to polish the OpenCL stuff...

Just to be clear - I haven't enabled the OpenCL stuff. I have a NVIDIA card, and I know madVR's OpenCL does not work on NVIDIA so I haven't bothered with it. This is simply ticking the box in 'Image Doubling' section.

Anyway, NVIDIA GTX 660Ti, Win 7 64bit.

[DragonQ]

Originally Posted by James Freeman

I disagree.
87.1b works much faster than even 86.11.
<snip>

You have Smooth Motion off (since you're playing a 30i file at 60 Hz). As I said a couple of posts ago, the performance regression is far worse with interlaced video when using Smooth Motion for me, so what happens when you enable it? Also, that's a ridiculous difference in performance, are you sure your settings haven't changed?

@madshi: 0.87.1b doesn't fix the issue.

When Smooth Motion is off the only difference between 0.86.11 and 0.87.1b is that "split" takes ~50% more time. With Smooth Motion on though, "split", "deinterlace", and "rendering" all take ~100% more time, and "present" takes ~200% more time.

[James Freeman]

Originally Posted by romulous

This is simply ticking the box in 'Image Doubling' section.

Image Doubling is new to v87 OpenCL feature.
So don't use it with NVIDIA (yet).

[madshi]

Originally Posted by James Freeman

I disagree.
87.1b works much faster than even 86.11.

Wow! I wish it would be this way for everybody...

Originally Posted by romulous

Just to be clear - I haven't enabled the OpenCL stuff. I have a NVIDIA card, and I know madVR's OpenCL does not work on NVIDIA so I haven't bothered with it. This is simply ticking the box in 'Image Doubling' section.

Anyway, NVIDIA GTX 660Ti, Win 7 64bit.

Image doubling / NNEDI3 does use OpenCL internally. I know, the settings dialog doesn't make it clear.

[Thunderbolt8]

DragonQ, your GT 430 is quite old by now, maybe thats where the big difference comes from.

[nevcairiel]

Originally Posted by James Freeman

I disagree.
87.1b works much faster than even 86.11.

These numbers look unrealistic. I seriously doubt their accuracy.

[romulous]

Originally Posted by madshi

Image doubling / NNEDI3 does use OpenCL internally. I know, the settings dialog doesn't make it clear.

Yes - the settings dialog is a mess, it certainly doesn't even hint at it, let alone make it clear.

Bad news - the 87.1b version breaks the previous fixed build. Fullscreening to second display is now broken again. This is without nnedi3 ticked of course.

[leeperry]

Well, I left "ITC Processing" enabled in the HD7850 drivers and that would supposedly do a lot of wonderful things on FS videos...maybe that's my problem, will run more tests.

Originally Posted by madshi

Have you tried native DXVA2 decoding instead?

Yep, either flavor of DXVA2 seems equally flaky and checking "use OpenCL to process DXVA2" in mVR doesn't help.

Originally Posted by madshi

As long as the CPU can handle the decoding load, what's wrong with using software decoding?

Well, CUVID worked beautifully IME and I planned on relying on hardware decoding for high bitrate H264/VC1 and I do run a bunch of VST plugins in ffdshow audio, but at the end of the day what matters is the real-world power consumption...I guess I'll have to run a bunch of tests using a wattmeter.

Originally Posted by madshi

However, OpenCL performance is better with AMD.

Yep, I've seen that and I think I really like the new error diffusion dithering

Originally Posted by madshi

If your GPU is fast enough then I would expect all queues to fill up nicely. That said, the whole OpenCL stuff in madVR is just a couple of days old, so there's very little experience how it affects the various queues.

Apparently, HWinfo32 is far more capable and accurate at showing the GPU/VRAM load than GPU-Z so I'll check it out and will get back to you on this one.

Originally Posted by nevcairiel

AMDs hardware decoder is rather bad compared to NVIDIA. Its extremely slow.

Fair enough, thanks for the detailed explanation.

Originally Posted by huhn

if you got a intel cpu try quicksync it's faster than nvidia dxva and worlds faster than amd dxva.

so I can use Intel's Quicksync for video decoding and output its result on the HD7850? Neato, will give it a shot.

[James Freeman]

Originally Posted by dragonq

you have smooth motion off (since you're playing a 30i file at 60 hz).

Originally Posted by nevcairiel

these numbers look unrealistic. I seriously doubt their accuracy.

Edit: I forgot to turn off Debanding in 87.1b, which means its even faster than whats in this image.

[Thunderbolt8]

testing the ayaka MSSL12.ts clip for performance (mpeg2, 1440x1080 upscaled to 1080p; 29,97 fps), 0.871b is a little faster for me than 0.8611

0.8611:
without deinterlacing: 17.20ms
with deinterlacing: 17.50ms + deint: 3.30ms


0.871b:
without deinterlacing: 16.00ms
with deinterlacing: 16.90ms + deint: 2.95ms

[huhn]

Originally Posted by madshi

Wow! I wish it would be this way for everybody...

these are all screens from 86.11



ivtc doesn't really care about queues but deinterlacing... how to judge performance this way.

even if amd in this case is changing the deinterlacing mode because the number of buffered frames is to low why are the rendertimes affected?

so I can use Intel's Quicksync for video decoding and output its result on the HD7850? Neato, will give it a shot.

you need a fake display or windows 8 for this. and your mainboard must have support for igpu. every mainboard with a vga dvi dp oder hdmi connector can do it and i'm not sure if all intel gpu can do quicksync all ix cpu for sure.

[hannes69]

Just tested 87.1b on AMD HD4550 (quite old, quite slow). Everything is working fine so far. I didn´t test everything (e.g. deinterlacing), but for my typical usage scenario everything works. No black or bluescreens, freezes, stuttering or other unwanted behaviour. I unticked every box in trade quality, ticked the opencl stuff. I tried NNEDI3 for chroma upscaling, it works. I only play 720p without scaling on a projector, by using software decoding, debanding (works great!), involving a 3DLUT and using Bicubic AR chroma upscaling without the need of smoothmotion (I have accurate refresh rates by custom Powerstrip resolutions) I get 70% GPU usage. So ALL features of recent madvr version concerning quality are still usable on my good old GPU. Clearly there is no more headroom left for playing back 1080p or scaling other than native resolution, but for my needs it´s great!
Great work madshi, I had to give a positive feedback right at this time, when 9 out of 10 users report problems at the moment.
Forget about NVIDIA, just define AMD or Intel as a system requirement for using madvr

[omarank]

Originally Posted by madshi

So the 871b build did not help at all compared to the official v0.87.1?

The performance is better in 871b as compared to the official v0.87.1. In the official version, I have performance issues with some progressive videos too which have been fixed in 871b.

Originally Posted by James Freeman

I disagree.
87.1b works much faster than even 86.11.

On my system, I am not able to see any performance improvement with 871b for interlaced videos.

[noee]

madshi:
With my VC1 1080i test file, .87b makes substantial perf gains for deint:

DXVA Native (FSE Mode):
.86: ~7.3ms, full queues, perfect playback
.871: ~18.2ms, starved render/backbuffer, many drops
.871b: ~4.10ms, full queues, perfect playback

DXVA CB (FSE Mode):
.871: ~7.5ms, full queues, perfect playback
.871b: ~2.75ms, full queues, perfect playback

HD6570

[DragonQ]

Originally Posted by Thunderbolt8

DragonQ, your GT 430 is quite old by now, maybe thats where the big difference comes from.

All of this testing is on my laptop using an Intel HD4000, as I've said several times (although I understand if you haven't been following the past dozen pages). I don't/can't use MadVR on my HTPC since MediaPortal doesn't support it.

One way to fix the problem is to use CUVID (on my NVS 5200M), since that handles the deinterlacing and then all MadVR has to do is mix frames and downscale on the HD4000 (75% load, same as with 0.86.11). I imagine the hugely increased stat times I'm seeing is due to the GPU being maxed out and thus not doing everything asked of it quickly enough when deinterlacing and smooth motion are both being used. The problem is likely a smaller performance regression that pushes the load from 85-90% (0.86.11) to above 100% (0.87.1b). I can't imagine the fact I have Optimus is an issue, since when the decoder is set to anything but CUVID the nVidia card has 0% load (and thus is presumably asleep).

[huhn]

Originally Posted by nevcairiel

These numbers look unrealistic. I seriously doubt their accuracy.

sorry but i got the same...

87.1b
normal(high) queue settings
low queue settings
for 86.11 see above

edit i have a watch at the hd4000 later maybe a intel only problem?