DGAVCDecNV 1.0.13: GPU decoding on Nvidia

[Guest]

I didn't mention that it plays fine using alpha 1 and it doesn't play at all using alpha 2 (tested with deinterlacing enabled and disabled). I accidentally had playback speed set to single step in the shipped INI file for alpha 2. Can you check that?

Best to delete the INI file and let a new one be created.

[Yoshiyuki Blade]

Glad to report that everything works fine on an 8400M GS, using Vista32.

[lucassp]

I accidentally had playback speed set to single step in the shipped INI file for alpha 2. Can you check that?

Best to delete the INI file and let a new one be created.

Thanks! It works fine now :)

[steaa]

Initally the application gave the "cuInt failed (100)" error when starting it up, and then when loading a file it gave the message "GPU decoder: Failed to create video decoder.". After copying the DLL files to System32 also, I had the same result.
I removed the dll files from System32 and then after installing CUDA driver and toolkit, it opened fine and loaded a file OK without any problem.


Nvidia 9600GT / Q6600 G0 Stepping / Windows XP (32Bit)

Results for BBC HD stream (1080i) -
Display enabled - 43 ~ 45fps. CPU - ~23%
Display disabled - 62 ~ 65fps. CPU - ~8%

Results for HD Suisse stream (720p) -
Display enabled - 71 ~ 74fps. CPU - ~23%
Display disabled - 106 ~ 113fps. CPU - ~7%

[Guest]

Thanks, guys.

DGAVCDecodeNV is well along. I expect to have an alpha in a day or two.

[Ranguvar]

Followed kemuri's instructions, but I still get artifacts with the non-lossless Touhou (link posted a while back), and the ther bugs mentioned in my first post. I also have the problem where eventually, it refuses to start, and I need to reboot.

Indexed with DGAVCIndex and decoded with DGAVCDecode: http://i37.tinypic.com/4qk6f8.png
Snapshot of the preview in DGAVCIndexNV (I had no way of playing the DGAVCIndexNV DGA, of course, so the flaw may only be in the preview): http://i36.tinypic.com/26449e1.png
I did not try DGAVCIndex's preview to see if it also has the flaw, as it does not let me seek to a specific frame, or even do anything to seek besides huge jumps and playback/pause.

Obviously, one uses different YV12 coefficients, but that's beside the point. Look at the bullets fired from the player. And no, this is not the lossless video, this is a "normal" H.264 stream. Other frames exhibit problems, too.

Alpha 2 used, w/o deinterlacing.

[Sagekilla]

Just would like to say, thanks for such a wonderful creation :) Greatly appreciate having hardware decode of H.264 for transcoding.

[Guest]

@Ranguvar

Please tell me what artifacts you are talking about. I looked at both your PNGs and I don't see anything. It's so frustrating when people just post a screenshot and go "see!". Make a circle around it or say it in words. I'm not a mind reader.

[Ranguvar]

There are huge differences between the correct DGAVCIndex one and the bad DGAVCIndexNV one.


DGAVCIndex (correct): http://i37.tinypic.com/4qk6f8.png
DGAVCIndexNV (modified to show artifact areas): http://i38.tinypic.com/2s7gkkj.png

[G_M_C]

Don't know if this is relevant or not; But does this new version have the same "DXVA"-limitations as other HW-assisted/DXVA decoders ?

i.e. Does it do L 5.1 x264's as well ?

[Zwitterion]

There are huge differences between the correct DGAVCIndex one and the bad DGAVCIndexNV one.


DGAVCIndex (correct): http://i37.tinypic.com/4qk6f8.png
DGAVCIndexNV (modified to show artifact areas): http://i38.tinypic.com/2s7gkkj.png
Probably the video can't be decoded by the GPU. I doubt that this is DGAVCIndexNV's fault. Try running it in MPC-HC and check the MPC Video Decoder whether DXVA is enabled or not.

It would be great if DGAVCIndexNV contained a check, so that people don't accidentally decode unsupported streams.

[CruNcher]

Yeah people create all kind of complexitys that are just to much for sane Hardware Decoders :) these block failures look exactly like the ones i experience with some X264 content and CoreAVC though might be not of the same nature but indeed ffdshow (libavcodec) doesn't show these kind of errors as it has a very primative error correction and can hide stream errors to some degress it seems (same for Mainconcepts Decoder) and these errors can also directly happen in the encoder not only because the stream has been damaged @ the transport. So i guess it would hide such occasional block errors very well from the viewer, CoreAVC can't do this but i guess Nvidias Hardware Decoder should also be able todo it like libavcodec does, as they expected it's being used for DVB watching :)

[Guest]

I see the artifacts now. Thank you.

I have a directory where I am collecting bad streams for CUDA. This has joined the collection. I'll be sending them to Nvidia for analysis. They've already fixed one bug exposed by some of my streams.

[Ranguvar]

Thanks very much :)

[AtomX]

So I realize that I'm being impatient, but I can't help but try and feed my curiosity. Has there been any progress on the decoder DLL? I'm really excited to start playing around with it.

[Guest]

Posted one day ago:

http://forum.doom9.org/showthread.php?p=1183558#post1183558

What do you want, hourly updates? I do have a life, you know. :)

It's basically working but there are some hangs on seeking that I am working on. If I don't sort it out in a few days I'll put out an alpha that at least does straight linear encode so that you can assess the effectiveness of the entire process when transcoding.

[kemuri-_9]

It's basically working but there are some hangs on seeking that I am working on. If I don't sort it out in a few days I'll put out an alpha that at least does straight linear encode so that you can assess the effectiveness of the entire process when transcoding.
sweet... looking forward to that at least.

[AtomX]

Posted one day ago:

http://forum.doom9.org/showthread.php?p=1183558#post1183558

What do you want, hourly updates? I do have a life, you know. :)

It's basically working but there are some hangs on seeking that I am working on. If I don't sort it out in a few days I'll put out an alpha that at least does straight linear encode so that you can assess the effectiveness of the entire process when transcoding.

Heh sorry to come off as a pest. As I said, I knew I was being impatient. Just excited to hear about progress, not trying to rush you at all.

[Guest]

I've just now found and fixed the problem with random access. I haven't coded anything for end of file yet so it just blows up. That shouldn't take me long. Looks good for an alpha tomorrow. The CUDA engine is a finicky beast and you have to get all your ducks in a row to reset it gracefully for a seek.

[Ranguvar]

Excellent! I remember that one. Does it also fix the error message pop-ups if you quit while playing?

[Guest]

Excellent! I remember that one. Does it also fix the error message pop-ups if you quit while playing? I don't know what you are talking about.

[Ranguvar]

2.) If I play either of those samples in the indexer, and then quit before stopping playback (with display on) I usually get the errors "cuvidMapVideoFrame: 201" and the same with the number 205, after I exit.

blahforcharacterlimit

[Guest]

That's in DGAVCIndexNV and I was talking about a fix to DGAVCDecodeNV. Still, I have managed now to duplicate it and I'll fix it. Thank you for reminding me about it.

[Guest]

Everything was looking good. I could open my indexed stream in VirtualDub, play and preview, and seek around.

Then I decided to try an encode. I fired up HCEnc and started an encoding. As the second pass began, HCEnc started reporting a mismatch in the number of frames between the first and second passes. Hmm, that's odd. And the resulting M2V was garbage.

Then I tried a simple HUFFYUV encode in VirtualDub. Worked great. Then a Cedocida DV encode. Worked great. Note that these are one-pass codecs.

Then I tried MeGUI with CE-Baseline. It crashed when I added the job to the queue. Then I tried DivX. It crashed.

So now I'm thinking, these guys are opening the AVS file twice or something, creating multiple instances of AVCSource(). So I opened the script in VirtualDub. Fine. Then I opened it again in VirtualDub, leaving the first one open too. Oops. The timeline looked just like the garbled M2V from HCEnc.

So the situation now is that I will try to verify my theory about the multiple opens and see if there is any mitigation for it (floating CUDA contexts?).

But be aware this has a strong potential to be a deal killer.

[kemuri-_9]

so there's a restriction that can only have one instance of the script open at once?

that shouldn't stop you from continuing to develop it, many encoders encode to lossless (ffvh, ffv1, lags, etc.) first
then encode the lossless in their lossy encoder of choice, so they wouldn't have problems using it.

[Sharktooth]

neuron2 cant you just fake the instances? i mean if the AVCSource() is using the same input, just redirect the second or nth instance to the first...

[Guest]

neuron2 cant you just fake the instances? i mean if the AVCSource() is using the same input, just redirect the second or nth instance to the first... I was thinking about that on the drive to work. I'll see if it can work. As long as the instances don't try to operate at the same time, it could be OK.

Meanwhile, do you know how MeGUI works inside? Am I correct that it opens the script several times? Why would it crash when I just add the job to the queue?

[Sharktooth]

yes, it opens several times coz before enqueuing the script is alredy open for displaying the preview. also megui stores 2 custom variables in the avs script for AR.
so when you hit enqueue the script is opened again for reading those vars and other stuff.
however the lack of multi-tasking (or multi-instancing) in cuda is quite a limiting factor.

[Guest]

yes, it opens several times coz before enqueuing the script is alredy open for displaying the preview. also megui stores 2 custom variables in the avs script for AR.
so when you hit enqueue the script is opened again for reading those vars and other stuff.
however the lack of multi-tasking (or multi-instancing) in cuda is quite a limiting factor. Ouch. CUDA can allow multi-instancing through floating contexts. I'll have to see if I can get it working. Every CUDA context switch would have to force the filter into seek mode.

After more thought, I don't think you can redirect an open filter instance to another one. I'd have to ask IanB about it, but I don't think Avisynth has that capability.

[Quark.Fusion]

How it will behave if another program uses CUDA? (i.e. F@H)

[Sagekilla]

Neuron2: Would you keep it still though, even if it completely bombs on 2-pass? There's still HUGE benefit for transcoding to an intermediary formats or for 1-pass CRF in x264.

I still would love to have a GPU accelerated decoder that can work for 1-pass CRF.


Edit: Any chance I could get my hands on that dll? ;)

[Guest]

Let me try it with 1-pass CRF and if it's OK I'll let you have it.

[Sagekilla]

Thanks! :) Greatly appreciate it.


Edit: neuron2, the behavior you describe effectively makes AVCSource() non-deterministic when it's called multiple times, correct?

[Quark.Fusion]

I think you can do two pass with that behaviour if you restart program before second pass.

[Guest]

I think you can do two pass with that behaviour if you restart program before second pass. Please pay attention. You cannot even queue the job in MeGUI.

[Guest]

x264 1-pass CRF encode works fine. I got 60 fps encode rate for 640x480 material. I don't normally do encodes, so I don't know if that is good.

I'll give it out this evening.

[crypto]

@neuron
Thanks for the update. I can't wait to check it out. Single instance should be ok for me.

[CruNcher]

thx neuron, that's great news these days 2pass encoding is anyways slowly dieing :)

[kemuri-_9]

thx neuron, that's great news these days 2pass encoding is anyways slowly dieing :)

uhhh no, plenty of people still use 2pass, it'll never die.

[Quark.Fusion]

Please pay attention. You cannot even queue the job in MeGUI.
Sorry, I'm using console, so just don't understand why it may fail in it (and I don't have build to test). Hope you will fix bug.

[Guest]

OK, I see. Sorry for my confusion. I'll try a 2-pass with x264 CLI tonight and let you know what happens.

[Sagekilla]

@kemuri-_9: That may be true, but crf is a very attractive alternative if you want a slightly faster encode (especially with prefiltering, or transcoding HD content) with consistent quality across the whole encode. Both are valid for what they're best in (2-pass for specific size, crf for specific quality).

[Guest]

Guys, please take 2-pass argument elsewhere.

[Guest]

Validated for x264 CLI one pass only:

http://neuron2.net/misc/DGAVCDecodeNV.zip

DO NOT post bug reports about this not working with MeGUI, etc.

This version will expire on 9-30-08. I hope to have a multi-instance version by then.

[kemuri-_9]

sweeeeeeeeeeeeeeeeeeeeeeeeeeeeeetttttttttttttttttt

[Guest]

Oops, redownload to get the users manual.

Set deinterlace=true for PureVideo deinterlacing. I'll be curious to see comparisons to Yadif, etc. It looks good to me, but remember a) I haven't used the fancy decombers a lot, and b) I'm a geezer; my eyes don't work good.

[rack04]

First of all, I want to personaly thank you for all your hard work. When I place the dll file into my avisynth plugin directory and load MeGUI avisynth script creator I get the following error message:

http://i11.photobucket.com/albums/a199/rack04/untitled-2.jpg

Now before you jump to conclusions, no I'm not trying to encode using the .dll. I'm not trying to create a script using the dll, though it appears that the dll file is automatically loaded.

[Guest]

http://neuron2.net/misc/NV.zip

Zip password is gpudecode

Put nvcuvid.dll somewhere in your DLL search path. System32 is good.

Preliminary tests indicate little significant speedup versus software decoding. Let's see what some other people achieve.

This is intended for testing with X264 CLI one-pass only. Do not put it in your Avisynth plugins directory. Do not try to use it with any GUI. Posts reporting issues with that will be ignored.

[Audionut]

Thanks. Your efforts are appreciated.

[Snake91]

System:
Windows XP SP3
nvidia 8800GT 1GB
ForceWare 177.98
Q6600 2.8Ghz

Playing speed: 43-60 fps (min-max), I tested on BD streams (Jumper and Transformers)
CPU usage: 4%

But when I try to save the project I get the usual cpu utilization (about 25% on only 1 core) and decoding speed

Thanks for your work