GPU Transcoding (Intel QuickSync and nVidia NVENC)

[Vicpa 611]

Hi @@babgvant,@Luke

 

Saw that ffmpeg 3.0 is out. Seems that Zeranoe's win build now includes the libmfx support. Is that build ok to use. (I am sure babgvant you would like to get out of the ffmpeg build task ​)

 

Was also wondering if the VPP problems were fixed.

 

thanks

vicpa

[babgvant 143]

AFAIK the VPP issues have been addressed. I can no longer reproduce here (last time I checked).

 

The build is scripted out so it doesn't matter to me either way .

[Vicpa 611]

Hi,

 

So you know what my next question will be.  Can the transcode of live tv be completed? Currently I understand the problem was with the encode side to h264.

 

With a few changes this is what I am getting with a HD  ​HDHR3-CC either

 

Stream #0:0 -> #0:0 (mpeg2video (mpeg2_qsv) -> h264 (libx264))
  Stream #0:1 -> #0:1 (ac3 (native) -> aac (native))

 

Stream #0:0 -> #0:0 (mpeg2video (mpeg2_qsv) -> h264 (libx264))
  Stream #0:1 -> #0:1 (copy)

 

I know Luke is not convinced that the copy is ready for primetime with hls. Which is fine. But can the libx264 be h264_qsv ? with the fix?

 

I realize it probably isn't quite that simple

 

Also I saw that the strict - experimental setting is no longer needed for aac.

 

 

Thanks

-vicpa

[Vicpa 611]

Hi,

 

Easy enough to try

i5 Edge client.

Hardware acceleration  "auto"

28-31% cpu utilization

transcode-68b43f16-f678-4b7b-923b-96d6493625c2.txt​

 

Hardware acceleration "qsv"

 

12-13% cpu utilization

 

transcode-acca5263-59ed-4702-83db-733a02fdcddc.txt

 

Very encouraging!!. (Awesome actually!!).. But being a realist what should I be looking for to make sure of the stability? Any special in the logs or just that it continues to work?

 

-Vicpa

[Vicpa 611]

Hi,

 

I guess there is always some caveat. -pix_fmt yuv420p  is not a valid format for h264_qsv. FFmpeg will change  it automatically to -pix_fmt nv12. I added a change to do it if qsv is selected., just to be correct. But now I am wondering how that will impact the client apps. Edge doesn't seem to care.

 

But is above my level of expertise for all the ramifications. Anyway I thought I would pass that on to whomever will look at this in an official capacity.

 

-vicpa

[babgvant 143]

Hi,

 

I guess there is always some caveat. -pix_fmt yuv420p  is not a valid format for h264_qsv. FFmpeg will change  it automatically to -pix_fmt nv12. I added a change to do it if qsv is selected., just to be correct. But now I am wondering how that will impact the client apps. Edge doesn't seem to care.

 

But is above my level of expertise for all the ramifications. Anyway I thought I would pass that on to whomever will look at this in an official capacity.

 

-vicpa

 

The pixel format used internal to ffmpeg won't matter to consumers. NV12 is a YUV format anyway, so it's more about binary format than conversion.

 

I wonder why Emby specifies pix_fmt...

[CFC 25]

@babgvant,  Any idea why Quicksync playback of .wtv recordings produce a 704x480 video output of 1280X720 content?

Here's the transcode log of a playback with quicksync :(QSV) transcode-8a7330fe-0a63-499d-837b-b14c6d8ca9df.txt

and here's the same file with it off: (Auto) transcode-ec468f70-83d5-4ef7-9557-6c44488c544e.txt

It works fine on live tv playback as well.

 

Thanks,

CFC

Edited March 3, 2016 by CFC

[babgvant 143]

@babgvant,  Any idea why Quicksync playback of .wtv recordings produce a 704x480 video output of 1280X720 content?

Here's the transcode log of a playback with quicksync :

and here's the same file with it off:

It works fine on live tv playback as well.

 

Thanks,

CFC

 

Short answer: WTV is a strange format.

 

Long answer: WTV writes 704x480 into the header metadata and expects the decoder to renegotiate the format (frame size) after playback begins. My guess is that the ffmpeg QSV decoder hasn't been made aware of this behavior.

[rcocchiararo 1]

do we, using emby stable on windows, need to download ffmpeg separately from zeranoe (or somewhere else) for something? (or at least, for quicksync?)

[Luke 42077]

no, you're all good for quicksync.

[CFC 25]

Short answer: WTV is a strange format.

 

Long answer: WTV writes 704x480 into the header metadata and expects the decoder to renegotiate the format (frame size) after playback begins. My guess is that the ffmpeg QSV decoder hasn't been made aware of this behavior.

Thanks.

So doesn't seem like there's anything to be done on the Emby side. Need update to ffmpeg decoder somehow.

Too bad.

Thanks again,

CFC

[MSattler 390]

Just out of curiosity, is anyone currently running something like an i7-4770 with quicksync that can show before or after?

 

To this point for over a year I've run virtualized, but am toying with going back to a physical server.  I need to weigh the ability to transcode more streams with the same hardware versus not being able to easily migrate a VM, back it up, etc.

 

 

Thanks!

[runtimesandbox 171]

Just out of curiosity, is anyone currently running something like an i7-4770 with quicksync that can show before or after?

 

To this point for over a year I've run virtualized, but am toying with going back to a physical server.  I need to weigh the ability to transcode more streams with the same hardware versus not being able to easily migrate a VM, back it up, etc.

 

 

Thanks!

 

This is exactly my dilemma. Virtualisation offers a number of Pro's but the downside is this. 

[MSattler 390]

This is exactly my dilemma. Virtualisation offers a number of Pro's but the downside is this. 

 

Yup.  Dell XPS 8900 right now in the outlet with coupon are $760 shipped for me.  i7-6700k, 16GB Ram, 2TB Disk w 32GB SSD Cache, and a GeForce GTX 745.  If that one box with GPU transcoding could outperform the 2 VM's I run right now on my i7-4770, and i7-3770 <one VM per host> then I would do it in a heartbeat.

[Cobra 6]

Yup.  Dell XPS 8900 right now in the outlet with coupon are $760 shipped for me.  i7-6700k, 16GB Ram, 2TB Disk w 32GB SSD Cache, and a GeForce GTX 745.  If that one box with GPU transcoding could outperform the 2 VM's I run right now on my i7-4770, and i7-3770 <one VM per host> then I would do it in a heartbeat.

 

 

I was under the impression QuickSync would work with emby running inside of a docker on unraid. I was planning a new unraid build with quicksync for this reason. I read further up in this thread that said docker should now work with quick sync.

[MSattler 390]

Docker probably works as its not a virtual machine emby runs on but a virtualized container. These are totally different.

[rcocchiararo 1]

Just out of curiosity, is anyone currently running something like an i7-4770 with quicksync that can show before or after?

 

To this point for over a year I've run virtualized, but am toying with going back to a physical server.  I need to weigh the ability to transcode more streams with the same hardware versus not being able to easily migrate a VM, back it up, etc.

 

 

Thanks!

 

quicksync makes transcoding 1080p and 720p content en a 2000 series dual core celeron viable, but in your case with an i7 (i have an i7 with no quicksync on my main pc), you will get less cpu stressing, and not much more (i guess).

[MSattler 390]

quicksync makes transcoding 1080p and 720p content en a 2000 series dual core celeron viable, but in your case with an i7 (i have an i7 with no quicksync on my main pc), you will get less cpu stressing, and not much more (i guess).

 

I get that overall % wise you would see higher increases on a dual core celeron.  That being said, I'm curious what happens with QuickSync if you send 3-4 streams at it, at the same time.   Perhaps that is the more important question... if QS is only good for 1-2 streams <for reducing stress on cpu>, then perhaps is not as big of a gain as expected for an i7.

[fredg89 8]

So I have a noob question. I want to know what's the difference between Quicksync and the Nvidia NVENC. Basically, if you set it to the NVENC, does the GPU work in tandem with the CPUs...or does it completely offload to the GPU? Reason I ask is because I'm trying to maximize trans coding potential.

[babgvant 143]

Thanks.

So doesn't seem like there's anything to be done on the Emby side. Need update to ffmpeg decoder somehow.

Too bad.

Thanks again,

CFC

 

Best bet is to open a ticket w/ the ffmpeg team.

[rcocchiararo 1]

So I have a noob question. I want to know what's the difference between Quicksync and the Nvidia NVENC. Basically, if you set it to the NVENC, does the GPU work in tandem with the CPUs...or does it completely offload to the GPU? Reason I ask is because I'm trying to maximize trans coding potential.

 

one uses intel's gpu, the other uses nvidias.

 

Then again, nvidia is not an option here.

[runtimesandbox 171]

one uses intel's gpu, the other uses nvidias.

 

Then again, nvidia is not an option here.

 

Is Nvidia NVENC ever going to be an option or is there something fundamental in emby / ffmpeg that means it won't work?

NVENC seems to be the only option for virtualised servers at the moment.

[eTomm 6]

Depends how you virtualize them If you use Xen you can use QS too

[runtimesandbox 171]

How are you virtualising? 

 

I use an i7-6600k on a Proxmox box. From my research it looks like it is potentially possible to pass through the intel GPU but its unstable and you loose any GUI access to other VM's?

 

 

Just re read your post. I was using Xen but switched over to proxmox before I decided to virtualise emby. 

Edited March 4, 2016 by spudy12

[eTomm 6]

Yes, I tested all the virtualization environments until I found the one that was allowing me to passthrough the intel GPU. I wasn't liking the idea to throw it unused. I use a skylake i7-6700 low power with a Q170 mb. I performed a ffmpeg test and I found a bug smilar to the bobgvant one (that should be corrected now If I didn't understand wrong)