madVR - high quality video renderer (GPU assisted)

[biggerpapi]

Originally Posted by Anime Viewer

Something else of note I noticed when I went in to add the forceVendor key and sting inside it I noticed there is a HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\Intel(R) HD Graphics 4000 key, but no sign of an Nvidia one.

I noticed the same thing. It seems madVR is trying to use the HD4000 for openCL. I even tried deleting the HD4000 value and added a new value for my nVidia gpu but as soon as I started a video it added the HD4000 right back.

[cyberbeing]

Originally Posted by Anime Viewer

I opened up regedit, went to HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL and created a new key named forceVendor, and in that created a string key which I named nVidia (you didn't mention it, but I'm guessing it should also have a value of 1), but it made no difference. I tried changing the name to Intel, and nothing ran any different.

I think you've misunderstood madshi's notation.

HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\forceVendor REG_SZ "nVidia"

This means to add a 'forceVendor' String (REG_SZ) with a value of 'nVidia' at HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL , not a key.

[Anime Viewer]

Originally Posted by cyberbeing

I think you've misunderstood madshi's notation.

HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\forceVendor REG_SZ "nVidia"

This means to add a 'forceVendor' String (REG_SZ) with a value of 'nVidia' at HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL , not a key.

Thanks for clearing that up. That worked, and now it has created a HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\GeForce GTX 680M key in the registry. 32 neurons seems to be the best setting for an Nvidia GTX 680M gpu chroma upscaling setting.

Its interesting...if I force the Nvidia NEEDI3 runs correctly, but if I force the Intel the screen goes green again...

[QBhd]

Originally Posted by seiyafan

Agreed, my 270x had OpenCL copy interop and kernel interop in the 400'ish fps, check your PCI-e

What OS do you use?

My R9 270x is only getting 185'ish fps

QB

[6233638]

Thanks for the fixes in this build - especially #097.
Now I can leave Smooth Motion enabled all the time.

[truexfan81]

Originally Posted by madshi

Just copy the settings.bin file. In order to go back to a stored settings.bin file, first run "restore default settings.bat", then copy the stored settings.bin file into the madVR folder. For any of this to work, madVR requires write access to its own folder. If you don't want to give madVR write access to its own folder, open regedit, browse to "HKCU\Software\madshi\madVR" and export to a reg file. Then when you later want to go back, again first run "restore default settings.bat", then double click the saved reg file.quite slow and results in very slow OpenCL <-> Direct3D9 interop performance.

good to know, thanks

[huhn]

Originally Posted by QBhd

What OS do you use?

My R9 270x is only getting 185'ish fps

QB

this has a lot to do with the cpu and pci-e versions.

i get over 440 fps with haswell and pci-e 3.0 ivy bridge is slower with pci-e 3.0.

just have a look at this post:
http://www.avsforum.com/t/1477339/so-youve-built-your-htpc-now-what-is-next-how-to-get-the-ultimate-picture-and-sound-quality-from-your-htpc-madvr-svp-xbmc-mediabrowser-jriver/450#post_24514253

this is not 100 % reproduceable number can varies between two identical cards and that a lot.

[nevcairiel]

Originally Posted by madshi

D'ya still have a Maxwell? Would be interesting to hear how that fares compared to the 660, considering that Maxwell seems to have noticeably improved compute performance over Kepler.

Originally Posted by ryrynz

Hey nevcairiel, any chance of that compute performance test of your 750?

Using the same settings as my GTX 660, it seems to produce similar performance - which is kinda impressive, considering its quite a bit slower in raw performance (~30%).

Chroma: Bicubic 75+AR
Image Up: Jinc3+AR
Image Down: CR+AR+LL

720->1080 with NNEDI 32N: 37ms with OD, 40ms with ED1. Stable and no dropped frames during a 24p movie.
If I step down image downscaling to CR without AR/LL, rendering goes down to ~32ms, or image upscaling down to lanczos3+AR to 35ms.

The hit from error diffusion is much lower than on the 660, but NNEDI does take a bit more performance out of it.
However, its quite usable with NNEDI for 24p content. A 750Ti, with around 20% more performance, should fare even better and beat the GTX 660 consistently. A "860" model based on Maxwell sounds like a good madVR card, but who knows when those will come out. To reach enough performance to run 60p with NNEDI, you'll need to wait for the new high end cards, though!

90% of all content I watch is 24p, 7% is 25p, and only the last 3% are anything higher, so for me it would probably even be enough.

[John Carmack]

Originally Posted by madshi

Try increasing the GPU queue size and the number of pre-presented frames for exclusive mode. Smooth motion FRC benefits from bigger queues.

I actually tried and noticed a lot of things:
-Disabling exclusive mode solves the stutter.
-When in exclusive mode, (even with a GPU queue of 24), the decoder queue is always something like 21-25/24 while the others are 20 or 21-24/24
-The mouse is stuttering a lot in exclusive mode (maybe a clue?)

[DragonQ]

Originally Posted by Asmodian

Wow! look at that interop fps, not good.

Do you have something else on the PCI-E bus? It looks like it takes a really long time for a frame to make a round trip.

Nope, no other PCI-E devices.

Originally Posted by seiyafan

Agreed, my 270x had OpenCL copy interop and kernel interop in the 400'ish fps, check your PCI-e

The only option in the BIOS that I can remember which relates to PCI-E sets the speeds of the 2nd and 3rd PCI-E x16 slots, which I'm not using. :/

The only other setting I can think of is the Windows Power Settings for PCI-E power savings but that makes no difference.

Originally Posted by Anime Viewer

Something else of note I noticed when I went in to add the forceVendor key and sting inside it I noticed there is a HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\Intel(R) HD Graphics 4000 key, but no sign of an Nvidia one.

Hmmmm, mine also only has a key for my old GTS 250 and "Pitcairn", whatever that is. There is no key for my R9 270...could this be the problem?

EDIT: Just realised Pitcairn is an AMD code name so that's probably fine.

[madshi]

Originally Posted by DragonQ

{benchmark image}

I don't think I'll be changing my GPU again, especially to nVidia out of principle.

Well, for whatever reason interop cost seems to be extremely high on your PC. How well does DXVA decoding with copyback work for you? I guess it'll probably be very slow, too. Many other users have better results. Could be a problem with your mainboard or your OS installation or something. Can't say for sure. At least it looks like you can forget about NNEDI3 with AMD on your mainboard or OS installation, unfortunately, unless AMD fixes the interop performance problem with a future driver.

Originally Posted by ryrynz

madshi, is there still reason to keep the old FSE path?

Some users are still using it. I hope to be able to get rid of it some day, but it's not costing me much support work, so I'll leave it in for now. The big cleanup might come for v1.0. No need to rush cleanup now.

Originally Posted by Anime Viewer

Thanks for clearing that up. That worked, and now it has created a HKEY_CURRENT_USER\Software\madshi\madVR\OpenCL\GeForce GTX 680M key in the registry. 32 neurons seems to be the best setting for an Nvidia GTX 680M gpu chroma upscaling setting.

Its interesting...if I force the Nvidia NEEDI3 runs correctly, but if I force the Intel the screen goes green again...

Ok, that's good to know. So NNEDI3 generally works on Optimus laptops. Now the big question is how I can automatically detect the situation, so that you don't have to create that registry tweak. I'll probably have to create a small test tool for you to run...

Originally Posted by 6233638

Thanks for the fixes in this build - especially #097.
Now I can leave Smooth Motion enabled all the time.

Originally Posted by nevcairiel

Using the same settings as my GTX 660, it seems to produce similar performance - which is kinda impressive, considering its quite a bit slower in raw performance (~30%).

Chroma: Bicubic 75+AR
Image Up: Jinc3+AR
Image Down: CR+AR+LL

720->1080 with NNEDI 32N: 37ms with OD, 40ms with ED1. Stable and no dropped frames during a 24p movie.
If I step down image downscaling to CR without AR/LL, rendering goes down to ~32ms, or image upscaling down to lanczos3+AR to 35ms.

The hit from error diffusion is much lower than on the 660, but NNEDI does take a bit more performance out of it.

That's quite interesting, thanks for the tests! FWIW, looking at raw numbers it seems that the 750 has about 55% of the memory bandwidth of the 660 and also about 55% of the GLOPS shader power. So it's quite promising if it runs at about the same speed in madVR.

You're saying the 750 is faster with error diffusion but a bit slower with NNEDI compared to the 660? I think the 750Ti should be a great HTPC card, at least it's by far the greatest card which doesn't need an extra power plug. Can't wait for 860Ti or 960Ti in 20nm without an extra power plug...

Originally Posted by John Carmack

I actually tried and noticed a lot of things:
-Disabling exclusive mode solves the stutter.
-When in exclusive mode, (even with a GPU queue of 24), the decoder queue is always something like 21-25/24 while the others are 20 or 21-24/24
-The mouse is stuttering a lot in exclusive mode (maybe a clue?)

The queues look just fine to me. Some questions:

(1) How's the presentation queue?
(2) Do you get frame drops? Or presentation glitches? Or both?
(3) Have you changed any other settings besides from the queue sizes?
(4) Hope you're using the default flush settings?
(5) Have you tried disabling desktop composition (if you're on win7)?

[TheShadowRunner]

Originally Posted by Mangix

Windows XP support ends on the 8th. I don't think it makes sense anymore.

Yeah well, 30% of the world thinks otherwise.
Please don't drop XP support.
And thanks madshi for the new build with Vsfilter/black frame bug fixed!

[DragonQ]

Originally Posted by madshi

Well, for whatever reason interop cost seems to be extremely high on your PC. How well does DXVA decoding with copyback work for you? I guess it'll probably be very slow, too. Many other users have better results. Could be a problem with your mainboard or your OS installation or something. Can't say for sure. At least it looks like you can forget about NNEDI3 with AMD on your mainboard or OS installation, unfortunately, unless AMD fixes the interop performance problem with a future driver.

OS is Windows 8.1 x64, couple of months old. Naturally all motherboard drivers are installed. Latest BIOS too.

Can't think of any settings to change so I guess I'm stuck. My CPU is overclocked but Nehalem uses Base Clock rather than CPU Multiplier for overclocking, so the PCI-E bus is overclocked as well. My RAM has to be underclocked due to this but I can't imagine that would make such a huge difference; it's still over 1333 MHz.

EDIT: Just realised the PCI-E bus isn't overclocked since it's separate to the Base Clock. Been a while since I've checked BIOS settings on this machine! In any case, QPI should surely be fast enough...is anyone else here using an AMD GPU on a QPI-based motherboard? I think QPI was only used on first-generation Core i7s (Bloomfield), at least for desktop systems.

[ryrynz]

Originally Posted by TheShadowRunner

Yeah well, 30% of the world thinks otherwise.

30% of the world won't upgrade unless made to also I'm sure 29.9% of them aren't using madVR. Anyway the old path is sticking around at least until 1.0 so don't worry.

[John Carmack]

Originally Posted by madshi

The queues look just fine to me. Some questions:

(1) How's the presentation queue?
(2) Do you get frame drops? Or presentation glitches? Or both?
(3) Have you changed any other settings besides from the queue sizes?
(4) Hope you're using the default flush settings?
(5) Have you tried disabling desktop composition (if you're on win7)

First, thank you for helping me, even if I'm an isolated case.

1) Almost always empty (i.e. 0 or 1-8/8)
2) Both
3) I tried with the reseted settings and it's the same.
4) Yes, defaults
5) Tried it too

I must precise again that these errors only show up with High10.

[Anime Viewer]

Originally Posted by madshi

Ok, that's good to know. So NNEDI3 generally works on Optimus laptops. Now the big question is how I can automatically detect the situation, so that you don't have to create that registry tweak.

While experimenting around with different forms of combinations it looks like optimus (and manybe AMD's hybrid) may be able to run in a hybrid state. While using the force registry key to Nvidia, and selecting the Intel gpu for MPC-HC in Nvidia control panel it looks like it may have used a combination of the two gpu. The NNEDI3 continued to render using the Nvidia because the screen didn't turn green and render times didn't shoot up drastically (like when I force the Intel in the registry), but the render times did increase ~5-7ms over the times when run with Nvidia control panel and the registry set to Nvidia. Potentially if other things besides OpenCL could be forced (like DirectCompute) people with dual gpu systems could eliminate the weaker of the two gpu's when it comes to using certain features. People with hybrid amd/intel systems might be able to get around the OpenCL interop bug if they set their systems to force the Intel in the registry, but tell their system to use the AMD for MPC-HC (or whatever other video player - like POT - they may be using).

[blu3wh0]

Originally Posted by DragonQ

I don't think I'll be changing my GPU again, especially to nVidia out of principle.

I would like to note that I'm getting the same issue as DragonQ with AMD 7950 x2. At first I thought I was crazy since most other people with AMD video cards seem to be getting at least some respectable luma doubling, thinking it might be a problem with my setup. I have tried everything that I could think of to try to fix this to no avail.

My OpenCL copy and kernel fps drops from approximately 5250 fps to 130 fps with the interop. I'm on Windows 8.1 x64 and running the latest AMD beta driver. OpenCl error diffusion used to give me the same problem before madshi switched it to DirectCompute. The lowest frame time I can get with luma doubling at 16 neurons for a 720p 24 fps file is 48-50 ms with everything else either disabled or the lowest possible settings. My GPU runs at approximately 40% when I try to use my current settings with luma doubling.

Sorry if this is kind of abrupt, I have been following madVR and lurking here since madshi started developing it, to which I cannot express my gratitude enough in words (for now, thank you!), but I have been pulling my hair out due to this problem. Please let me know if I can provide any other kind of info.

[flashmozzg]

Originally Posted by blu3wh0

I would like to note that I'm getting the same issue as DragonQ with AMD 7950 x2. At first I thought I was crazy since most other people with AMD video cards seem to be getting at least some respectable luma doubling, thinking it might be a problem with my setup. I have tried everything that I could think of to try to fix this to no avail.

My OpenCL copy and kernel fps drops from approximately 5250 fps to 130 fps with the interop. I'm on Windows 8.1 x64 and running the latest AMD beta driver. OpenCl error diffusion used to give me the same problem before madshi switched it to DirectCompute. The lowest frame time I can get with luma doubling at 16 neurons for a 720p 24 fps file is 48-50 ms with everything else either disabled or the lowest possible settings. My GPU runs at approximately 40% when I try to use my current settings with luma doubling.

Sorry if this is kind of abrupt, I have been following madVR and lurking here since madshi started developing it, to which I cannot express my gratitude enough in words (for now, thank you!), but I have been pulling my hair out due to this problem. Please let me know if I can provide any other kind of info.

What cpu / pci-e do you have?

[blu3wh0]

I have an i5-2500k on an ASRock Z67 Extreme4 motherboard, with the video cards running on PCI Express 2.0 x8/x8. Disabling crossfire also provided no improvements.

[huhn]

pci-e 2.0 x8 is like using pci-e 1.1 and this is to slow for nnedi on amd. i hope amd fix this in the future.