madVR - high quality video renderer (GPU assisted)

[har3inger]

I've gone ahead and done that. I tested:

D3D11 old, 3 present queue: Everything's fine.
D3D9, 3 present queue: Everything's fine.
D3D 11 new, 3 present queue: Everything's fine.

Here's where it's interesting:
D3D9, 16 present queue: a/v desync. Video is noticeably earlier than audio, but you'd have to be looking for it to see it. Didn't bother checking if it gets better/worse over time. I feel like it's a constant displacement.
D3D11 New, 15 present queue (setting it to 16 makes it 15): a/v desync. Exact same amount of perceived desync as D3D9, with video noticeably earlier than audio.
D3D11 Old, 15 present queue: Never fills the present queue above 4-5/15, never fills GPU queue above 3-4/24. Oddly enough, starts with mad stuttering and tons of dropped frames, but after queues stabilize, it settles down and stops dropping frames, with no A/V desync.

All of these above tests are using FSE mode, and I quit the program and reload after each setting just to be sure they're getting applied properly. The OSD reports that the present queues and D3D mode are set as I wanted. Oh yeah, D3D9 isn't throwing tons of glitches anymore. Everything is 0 glitches, 0 dropped frames after queues fill.

Now, in windowed, playing with backbuffer sizes:
D3D9: backbuffer 16, present queue 16: A/V desync in FSE, back out into windowed mode, desync is gone.
D3D11 new: present queue (windowed) 15, present queue FSE 15: A/V desync in FSE, gone in windowed.

It's definitely a problem with the present queues in FSE, and isn't a result of whether I use D3D11 or D3D9, since it affects both. The problem exists in FSE, but disappears in windowed mode with the same number of backbuffer/present frames. I suspect there's desync as well with smaller present queues, but it's been unnoticeable since I set it to max of 3.

[Werewolfy]

Originally Posted by madshi

Glad to hear that! Looking forward to your screenshots.

Here is a screenshot that shows clearly the difference between Linear Light on and off for Crispen Edges.

http://screenshotcomparison.com/comparison/164352

The value I used was 1.0 in upscaling refinement but the issue is also present in image enhancements. I guess it's ringing after all but it's pretty bad...

[XMonarchY]

Does "Use Alternative Glitch-Handling Mode" option reduce visual quality and/or performance in any way?

[AngelGraves13]

Originally Posted by XMonarchY

Does "Use Alternative Glitch-Handling Mode" option reduce visual quality and/or performance in any way?

Nope. While I'm not too sure what it does exactly, it does seem to be something along the lines of "alternate vsync" basically having less glitches than regular vsync.

[madshi]

Originally Posted by har3inger

I've gone ahead and done that. I tested:

D3D11 old, 3 present queue: Everything's fine.
D3D9, 3 present queue: Everything's fine.
D3D 11 new, 3 present queue: Everything's fine.

Here's where it's interesting:
D3D9, 16 present queue: a/v desync. Video is noticeably earlier than audio, but you'd have to be looking for it to see it. Didn't bother checking if it gets better/worse over time. I feel like it's a constant displacement.
D3D11 New, 15 present queue (setting it to 16 makes it 15): a/v desync. Exact same amount of perceived desync as D3D9, with video noticeably earlier than audio.
D3D11 Old, 15 present queue: Never fills the present queue above 4-5/15, never fills GPU queue above 3-4/24. Oddly enough, starts with mad stuttering and tons of dropped frames, but after queues stabilize, it settles down and stops dropping frames, with no A/V desync.

All of these above tests are using FSE mode, and I quit the program and reload after each setting just to be sure they're getting applied properly. The OSD reports that the present queues and D3D mode are set as I wanted. Oh yeah, D3D9 isn't throwing tons of glitches anymore. Everything is 0 glitches, 0 dropped frames after queues fill.

Now, in windowed, playing with backbuffer sizes:
D3D9: backbuffer 16, present queue 16: A/V desync in FSE, back out into windowed mode, desync is gone.
D3D11 new: present queue (windowed) 15, present queue FSE 15: A/V desync in FSE, gone in windowed.

It's definitely a problem with the present queues in FSE, and isn't a result of whether I use D3D11 or D3D9, since it affects both. The problem exists in FSE, but disappears in windowed mode with the same number of backbuffer/present frames. I suspect there's desync as well with smaller present queues, but it's been unnoticeable since I set it to max of 3.

Ok, that means the new mode isn't worse than the old, which is good.

Can you roughly quantify the desync amount?

In any case, it kinda looks like a driver issue. My impression is that for some reason your driver doesn't allow a higher than 4 frame queue. The "old" D3D11 mode just had different problems with that than the "new" mode. I assume you have tried different drivers already, and also a "clean" driver install with default settings etc?

Found this:

http://forums.guru3d.com/showthread.php?t=403389

So it seems there is a registry setting for the flip queue size for AMD drivers. Maybe it's set to a non-standard value for your GPU for some reason? Try setting it to 16 as a test.

Originally Posted by Werewolfy

Here is a screenshot that shows clearly the difference between Linear Light on and off for Crispen Edges.

http://screenshotcomparison.com/comparison/164352

The value I used was 1.0 in upscaling refinement but the issue is also present in image enhancements. I guess it's ringing after all but it's pretty bad...

Uh, that does look ugly!

Can you please make the original unsharpened frame available (PNG should suffice), so that I can test with that?

[AngelGraves13]

If your queue isn't filling up, you're probably overtaxing the GPU. Test and see what happens when you set all the scaling to bilinear for example and don't use doubling. Does it still not fill up? Then...as madshi said, it's a driver issue.

[Unr3aL]

Originally Posted by aufkrawall

You probably rather downloaded the same picture twice.
I checked the image for different pixels (I downloaded them from the image hoster as well), each red pixel differs:

You are probably right. After downloading the two a third time (already did twice before) it finally showed the difference... Maybe some firefox caching problem or something along the lines...

Sorry though.

Greetz, Unr3aL

[Werewolfy]

Originally Posted by madshi

Uh, that does look ugly!

Can you please make the original unsharpened frame available (PNG should suffice), so that I can test with that?

Sure, here you go

Original

[har3inger]

Originally Posted by AngelGraves13

If your queue isn't filling up, you're probably overtaxing the GPU. Test and see what happens when you set all the scaling to bilinear for example and don't use doubling. Does it still not fill up? Then...as madshi said, it's a driver issue.

They "fill" from zero and get stuck. It's definitely not dropping queues, which is what it would look like if GPU isn't keeping up. Render times can be <5ms and it wouldn't matter.

The A/V delay is probably on the order of 25-75 ms. It's slight, but it's there. Has anyone else noticed this problem? Only exists when you use very high present queues in FSE.

I've tried only 16.2 crimson for testing the A/V delay. The driver is kinda cruddy and has lots of broken UI/settings functions on laptop, but otherwise runs applications well. Things are all set at default. I don't really want to revert to 15.12 crimson because the new driver is needed for a couple games I have.

Radeonpro doesn't actually work on my computer. It can't find the ATI card. Regedit search shows that the reg key doesn't exist by default. I'll go add it in manually at some point later once I find where my GPU reg settings are.

For now, I might just go back to using a 3 frame queue for present queue. Honestly, I don't really see any difference between 3 and something huge like 12, and FRC behavior doesn't seem to change based on the queue size. I've read before that you need more frames in present queue for FRC, but haven't actually found this to be the case. What would I be looking for, if it's not working properly at low queue sizes?

[AngelGraves13]

Are you limiting your frames to 3 or less in your AMD control panel? Does it ONLY affect FSE?

[har3inger]

Only affects FSE. It's not possible to limit frames to anything in AMD control panel. This problem with weird queues is the entire reason madshi rolled out a new fix for the d3d11 render code.

[madshi]

Originally Posted by Werewolfy

Sure, here you go

Original

Thanks!

Originally Posted by har3inger

The A/V delay is probably on the order of 25-75 ms. It's slight, but it's there. Has anyone else noticed this problem? Only exists when you use very high present queues in FSE.

Weird. If the queue only fills to 4 frames with the "old" D3D11 mode, then a desync of maybe 15-4 = 11 frames would have been logical.

Originally Posted by har3inger

Radeonpro doesn't actually work on my computer. It can't find the ATI card.

There are a bunch of ATI tweak tools, this was just one which had a nice picture to it, showing the flip queue size GUI element. I suppose most/all of them will have the setting, so you could try a different tool. But if the desync is that small, I've kinda changed my mind and doesn't really expect this tweak to produce a difference. If you had a delay of ~11 frames then it would be a different situation.

Originally Posted by har3inger

For now, I might just go back to using a 3 frame queue for present queue. Honestly, I don't really see any difference between 3 and something huge like 12, and FRC behavior doesn't seem to change based on the queue size. I've read before that you need more frames in present queue for FRC, but haven't actually found this to be the case. What would I be looking for, if it's not working properly at low queue sizes?

If you neither get frame drops nor presentation glitches, and if everything looks smooth to your eyes, too, then all is fine and you can just continue using the small queue size.

Originally Posted by har3inger

This problem with weird queues is the entire reason madshi rolled out a new fix for the d3d11 render code.

The original reason I implemented that was to fix shaky 3D playback. But after I implemented it, I thought it might help with non-filling presentation queues, too.

[Werewolfy]

I've found a really weird issue, I don't know if it has been reported.

The queues are not filling (all of them) when I upscale a video in 2160p using D3D11 exclusive mode (10 bit).
It's the only situation when this issue happens. I have no issues with D3D11 windowed mode or with D3D11 exclusive mode (8bit) or when I select 1080p for the desktop (whatever the resolution of the movie).

I was using Windows 7 x64 and never saw this issue. I've noticed it since I installed Windows 8.1 x64.

[madshi]

Originally Posted by Werewolfy

I've found a really weird issue, I don't know if it has been reported.

The queues are not filling (all of them) when I upscale a video in 2160p using D3D11 exclusive mode (10 bit).
It's the only situation when this issue happens. I have no issues with D3D11 windowed mode or with D3D11 exclusive mode (8bit) or when I select 1080p for the desktop (whatever the resolution of the movie).

I was using Windows 7 x64 and never saw this issue. I've noticed it since I installed Windows 8.1 x64.

That has been reported a couple of times. The usual workaround in the past has been to reduce the number of prepresented frames to 3 or 4 frames. But please try the new "alternative glitch handling mode", maybe that fixes the issue for you? If it does, please double check that you don't get audio/video desync issues. I don't think you should, but since it's a brand new feature, double checking that might make sense.

[6233638]

Originally Posted by madshi

That has been reported a couple of times. The usual workaround in the past has been to reduce the number of prepresented frames to 3 or 4 frames. But please try the new "alternative glitch handling mode", maybe that fixes the issue for you? If it does, please double check that you don't get audio/video desync issues. I don't think you should, but since it's a brand new feature, double checking that might make sense.

For what it's worth, the new glitch handling mode didn't help for me with a GTX 960. I still have to keep my present queue set to 6.

[Werewolfy]

Originally Posted by madshi

That has been reported a couple of times. The usual workaround in the past has been to reduce the number of prepresented frames to 3 or 4 frames. But please try the new "alternative glitch handling mode", maybe that fixes the issue for you? If it does, please double check that you don't get audio/video desync issues. I don't think you should, but since it's a brand new feature, double checking that might make sense.

I tried the "alternative glitch handling mode" but it didn't really help. Sometimes when I move forward or backward (I don't mean when displyaing the seekbar) into the video the queues filled but it's pretty rare.

If I reduce the prepresented frames to 4 frames, the queues fill correctly but I don't really want to modify that... The only combination that doesn't give me dropped frames with interlaced DVD are :
- CPU queue size : 23
- GPU queue size : 14
- Frames presented in advance : 6

Deinterlacing seems to be pretty erratic most of the time.

Originally Posted by 6233638

For what it's worth, the new glitch handling mode didn't help for me with a GTX 960. I still have to keep my present queue set to 6.

Strange, 6 doesn't work for me. I have to use 4.

[har3inger]

Okay, this is weird. I tried turning off FSE and just using windowed FS for longer queues, and D3D11 windowed full screen has a delay, but windowed (downsized 1080p-> 1080p plus window frames) has no delay.

Whatever. Back to tiny queue sizes it is :P

[GCRaistlin]

Jump to time (by clicking on the time scale in MPC-HC) is broken since 0.90.4 (WinXP). To be precise, the first click works but the second one is not and the player may hang after it.
Sample: http://mir.cr/2XNEHWTK

[CarlosCaco]

with d3d11 "old" my queues don´t fill, with the new alternate now my queues fill and the a/v are perfect synced, but i have a slight perfomance loss because the present a frame for every vsync option

i don´t output 10 bit and 3d, so there any advantage for me to use d3d11 ? or should i stay on d3d9?

i´m usinf FSE with the standard settings:

- CPU queue size : 16
- GPU queue size : 8
- Frames presented in advance :8

theres any advantage if i increase my queues? i have the trade quality option don´t render frames when fade... checked

[ryrynz]

Originally Posted by 6233638

For what it's worth, the new glitch handling mode didn't help for me with a GTX 960. I still have to keep my present queue set to 6.

I have a 960 also. For what content and settings? Upscaling to 4k using d3d11 10-bit?