[Announce] VA-API/VPP Patch for vdr-plugin-softhddevice - v10

rofafor · November 27, 2014 at 7:45 PM

Hi, a new version of VAAPI patch is now available for testing. On my Haswell setup the picture quality is much better than in earlier versions and even MCDI is finally usable. Also, the stuttering on Burosch samples is now neglible.
Please, remember to reset both FieldHistory values to 0 before reporting any feedback (in German or English) on this thread! We are hoping to get rid of those field order settings in the setup menu.

Quote

This is a VA-API VPP patch for softhddevice. Some changes over the previous version:

* Fix postprocessing reference surface ordering bug. Hopefully FieldHistory -values can now be set to 0|0
* Massive performance boost to postprocessing (5-10x)
* Fix black surface creation
* Improvements in surface synchronization
* Add support for picture sharpening via gpe pipeline (warning: sharpening may be buggy in libva-intel-driver)
* Add HW-accelerated support for GRAB

This patch adds support for Intel VA-API VPP filters to softhddevice. It is expected that many
things that used to work are now broken (different frame rates, etc.) or that the entire patch
does not work well enough for production use.

This patch allows configuring the denoise, sharpening and deinterlace mode via plugin menu. In addition
there is some support for changing picture controls (brightness & contrast) from the plugin menu.

It is strongly recommended to use git master version of libva-intel-driver together with this patch
(http://cgit.freedesktop.org/vaapi/intel-driver/?h=master). Minimum recommended released version is
libva-intel-driver 1.4.1 which contains necessary fixes for the patch to function properly.

Some additional TODO items for the patch:
* Providing the picture control limits to the plugin menu (e.g. hue range from -180 to 180)
* Error handling in case certain vaapi features are not available
* Testing / fixing various framerate and display refreshrate combinations

The patch was/is tested and developed on:
* ffmpeg 1.2.6
* libva 1.4.1
* libva-intel-driver 1.4.2.pre1
* xf86-video-intel 2.99.916
+ Intel D54250WYK NUC kit (Haswell i5)

Display More

Download: 0001-VA-API-VPP-support-to-softhddevice-v10.patch

Example of required command line parameters:

Code

~# vdr -P 'softhddevice -a hw:0,3 -d :0.0 -f -v va-api'

One can check the current GPU load with the intel_gpu_top application from x11-apps/intel-gpu-tools package:

Code

~# emerge x11-apps/intel-gpu-tools # using Gentoo
~# intel_gpu_top

All credits go to Antti Seppälä for his superb work!

Copperhead · November 27, 2014 at 9:06 PM

MotionCompensated deinterlacing doesn't seem to work at all. Massive interlace artifacts in fast moving objects.

MotionAdaptive deinterlacing works quite good with first field: 2 and second field 1. But if you take a closer look at moving texts you can spot some juddering.

libva : git master
libva-intel-driver: git master
ffmpeg: 2.4.3
xf86-video-intel: 2.99.916
+ Intel Celeron J1900

rofafor · November 27, 2014 at 9:25 PM

MCDI requires gen 7.5 or 8 hardware (according to http://cgit.freedesktop.org/vaapi/intel-dr…5_device_info.c) and your Celeron seems to be gen 7. However, Gwenole's repo contain MCDI code for gen 7 hardware.

Please, try this driver: https://github.com/gbeauchesne/li…tree/19.vpp.adi

3PO · November 27, 2014 at 9:45 PM

How can i find out, if my cpu is supported?

fnu · November 27, 2014 at 10:00 PM

Quote from rofafor

All credits go to Antti Seppälä for his superb work!

Yes, this must be marked loud and clear, thanks for that passion and patience

Will test on BayTrail and SandyBridge and come back ... MADI w/o micro stutter would be great. Maybe I find some HW to test MCDI, but I suspect the difference is more or less subtle ...

Regards
fnu

fnu · November 27, 2014 at 10:08 PM

Quote from Copperhead

MotionCompensated deinterlacing doesn't seem to work at all. Massive interlace artifacts in fast moving objects.

Have you set both FieldHistory to "0" for MCDI?

I suspect the little BayTrail might lack a bit performance for MCDI, but doesn't matter, we get 1080i on temporal level at 14W ...

Regards
fnu

Copperhead · November 27, 2014 at 10:14 PM

Quote from fnu

Have you set both FieldHistory to "0" for MCDI?

I tried every possible combination of these two settings. No difference.

Quote from rofafor

Please, try this driver: https://github.com/gbeauchesne/libva-int…tree/19.vpp.adi

Thank you for the link. I'm going to test it tomorrow.

Ulrich Eckhardt · November 29, 2014 at 7:05 PM

I tried this on my new ASRock Q2900M, but the VDR crashes. The still picture from the radio-plugin works, but when switching to ARD it crashes. This is my first attempt to set up a VDR with intel drivers so I am not sure if all my versions are correct.

The famous last words in the log:

Code

Nov 29 18:42:50 vdr-lan vdr[17943]: video: --:--:--.---   +0    0   0/\ms   0+0 v-buf
Nov 29 18:42:50 vdr-lan vdr[17943]: video: decoder buffer empty, duping frame (8843/4) 0 v-bu
Nov 29 18:42:50 vdr-lan vdr[17943]: video/vdpau: missed frame (6495/4)
Nov 29 18:42:50 vdr-lan vdr[17943]: video: decoder buffer empty, duping frame (8844/4) 0 v-bu
Nov 29 18:42:50 vdr-lan vdr[17943]: video: --:--:--.---   +0    0   0/\ms   0+0 v-buf
Nov 29 18:42:50 vdr-lan vdr[17943]: codec/video: ffmpeg/libav buggy: width or height zero

I compiled the VDR and intel drivers by myself on an OpenSUSE 13.2
vainfo:

Code

vdr-lan:~ # vainfo 
libva info: VA-API version 0.36.0
libva info: va_getDriverName() returns 0
libva info: Trying to open /usr/lib64/dri/i965_drv_video.so
libva info: Found init function __vaDriverInit_0_36
libva info: va_openDriver() returns 0
vainfo: VA-API version: 0.36 (libva 1.4.2.pre1)
vainfo: Driver version: Intel i965 driver for Intel(R) Bay Trail - 1.4.2.pre1 (1.3.2-97-gedbb8cd)
vainfo: Supported profile and entrypoints
      VAProfileMPEG2Simple            : VAEntrypointVLD
      VAProfileMPEG2Simple            : VAEntrypointEncSlice
      VAProfileMPEG2Main              : VAEntrypointVLD
      VAProfileMPEG2Main              : VAEntrypointEncSlice
      VAProfileH264ConstrainedBaseline: VAEntrypointVLD
      VAProfileH264ConstrainedBaseline: VAEntrypointEncSlice
      VAProfileH264Main               : VAEntrypointVLD
      VAProfileH264Main               : VAEntrypointEncSlice
      VAProfileH264High               : VAEntrypointVLD
      VAProfileH264High               : VAEntrypointEncSlice
      VAProfileH264StereoHigh         : VAEntrypointVLD
      VAProfileVC1Simple              : VAEntrypointVLD
      VAProfileVC1Main                : VAEntrypointVLD
      VAProfileVC1Advanced            : VAEntrypointVLD
      VAProfileNone                   : VAEntrypointVideoProc
      VAProfileJPEGBaseline           : VAEntrypointVLD

Display More

ffmpeg

Code

vdr-lan:/video/src/va/intel-driver # ffmpeg 
ffmpeg version 2.3.3 Copyright (c) 2000-2014 the FFmpeg developers
  built on Nov 14 2014 12:17:00 with gcc 4.8 (SUSE Linux)

snippet from xorg.0.log

Code

[  1520.070] (II) xfree86: Adding drm device (/dev/dri/card0)
[  1520.072] (--) PCI:*(0:0:2:0) 8086:0f31:1849:0f31 rev 14, Mem @ 0xd0000000/4194304, 0xc000
0000/268435456, I/O @ 0x0000f080/8
[  1520.072] (--) PCI: (0:1:0:0) 18c3:0720:18c3:dd00 rev 1, Mem @ 0xd0710000/65536, 0xd070000
0/65536
[  1520.072] (II) LoadModule: "glx"
[  1520.072] (II) Loading /usr/lib64/xorg/modules/extensions/libglx.so
[  1520.074] (II) Module glx: vendor="X.Org Foundation"
[  1520.074]    compiled for 1.16.1, module version = 1.0.0
[  1520.074]    ABI class: X.Org Server Extension, version 8.0
[  1520.074] (==) AIGLX enabled
[  1520.074] (II) LoadModule: "intel"
[  1520.074] (II) Loading /usr/lib64/xorg/modules/drivers/intel_drv.so
[  1520.075] (II) Module intel: vendor="X.Org Foundation"
[  1520.075]    compiled for 1.16.1, module version = 2.99.916
[  1520.075]    Module class: X.Org Video Driver

Display More

And a backtrace:

Code

#0  0x00007f213e1af187 in raise () at /lib64/libc.so.6
#1  0x00007f213e1b0538 in abort () at /lib64/libc.so.6
#2  0x00007f213e1a8126 in __assert_fail_base () at /lib64/libc.so.6
#3  0x00007f213e1a81d2 in  () at /lib64/libc.so.6
#4  0x00007f2116dc059b in intel_decoder_sanity_check_input (decode_state=0x171cb08, h264_profile=24234760, ctx=0x171a6c0) at i965_decoder_utils.c:846
#5  0x00007f2116dc059b in intel_decoder_sanity_check_input (ctx=ctx@entry=0x171a6c0, profile=profile@entry=VAProfileH264Main, decode_state=decode_state@entry=0x171cb08)
    at i965_decoder_utils.c:1088
#6  0x00007f2116d9cd6c in gen7_mfd_decode_picture (ctx=0x171a6c0, profile=VAProfileH264Main, codec_state=0x171cb08, hw_context=0x7f20d421d600) at gen7_mfd.c:2615
#7  0x00007f2139705e8f in vaEndPicture (dpy=0x1794b60, context=33554432) at va.c:1129
#8  0x00007f212c395dde in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1
#9  0x00007f212c39657b in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1
#10 0x00007f212c39877b in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1
#11 0x00007f213c3b1958 in VideoDrawRenderState (hw_decoder=0x1ece0c0, vrs=0x7f20d4607460)
    at video.c:10182
#12 0x00007f213c3b8760 in Codec_draw_horiz_band (video_ctx=0x7f20d42d6f60, frame=0x7f20d42f6d20, offset=0x7f20e2b50b80, y=0, type=3, height=720) at codec.c:360
#13 0x00007f213a696698 in  () at /usr/lib64/libavcodec.so.55
#14 0x00007f213a9f573b in  () at /usr/lib64/libavcodec.so.55
#15 0x00007f213a6d8838 in  () at /usr/lib64/libavcodec.so.55
#16 0x00007f213a699dc8 in  () at /usr/lib64/libavcodec.so.55
#17 0x00007f213a9b378e in avcodec_decode_video2 () at /usr/lib64/libavcodec.so.55
#18 0x00007f213c3b8c7b in CodecVideoDecode (decoder=0x1089f70, avpkt=0x7f213c5c8d10 <MyVideoStream+2000>) at codec.c:604
#19 0x00007f213c3a7f7c in VideoDecodeInput (stream=0x7f213c5c8540 <MyVideoStream>)
    at softhddev.c:1965
#20 0x00007f213c3b07e1 in VdpauDisplayHandlerThread () at video.c:9033
#21 0x00007f213c3b1561 in VideoDisplayHandlerThread (dummy=0x0) at video.c:9914
#22 0x00007f213f8360a4 in start_thread () at /lib64/libpthread.so.0
#23 0x00007f213e25f7fd in clone () at /lib64/libc.so.6
(gdb) bt
#0  0x00007f213e1af187 in raise () at /lib64/libc.so.6                                          
#1  0x00007f213e1b0538 in abort () at /lib64/libc.so.6                                           
#2  0x00007f213e1a8126 in __assert_fail_base () at /lib64/libc.so.6                                
#3  0x00007f213e1a81d2 in  () at /lib64/libc.so.6                                                   
#4  0x00007f2116dc059b in intel_decoder_sanity_check_input (decode_state=0x171cb08, h264_profile=24234760, ctx=0x171a6c0) at i965_decoder_utils.c:846
#5  0x00007f2116dc059b in intel_decoder_sanity_check_input (ctx=ctx@entry=0x171a6c0, profile=profile@entry=VAProfileH264Main, decode_state=decode_state@entry=0x171cb08)
    at i965_decoder_utils.c:1088                                                                          
#6  0x00007f2116d9cd6c in gen7_mfd_decode_picture (ctx=0x171a6c0, profile=VAProfileH264Main, codec_state=0x171cb08, hw_context=0x7f20d421d600) at gen7_mfd.c:2615
#7  0x00007f2139705e8f in vaEndPicture (dpy=0x1794b60, context=33554432) at va.c:1129                       
#8  0x00007f212c395dde in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1                                        
#9  0x00007f212c39657b in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1                                         
#10 0x00007f212c39877b in  () at /usr/lib64/vdpau/libvdpau_va_gl.so.1                                           
#11 0x00007f213c3b1958 in VideoDrawRenderState (hw_decoder=0x1ece0c0, vrs=0x7f20d4607460)                       
    at video.c:10182                                                                                            
#12 0x00007f213c3b8760 in Codec_draw_horiz_band (video_ctx=0x7f20d42d6f60, frame=0x7f20d42f6d20, offset=0x7f20e2b50b80, y=0, type=3, height=720) at codec.c:360
#13 0x00007f213a696698 in  () at /usr/lib64/libavcodec.so.55                                                                                                 
#14 0x00007f213a9f573b in  () at /usr/lib64/libavcodec.so.55
#15 0x00007f213a6d8838 in  () at /usr/lib64/libavcodec.so.55
#16 0x00007f213a699dc8 in  () at /usr/lib64/libavcodec.so.55
#17 0x00007f213a9b378e in avcodec_decode_video2 () at /usr/lib64/libavcodec.so.55
#18 0x00007f213c3b8c7b in CodecVideoDecode (decoder=0x1089f70, avpkt=0x7f213c5c8d10 <MyVideoStream+2000>) at codec.c:604
#19 0x00007f213c3a7f7c in VideoDecodeInput (stream=0x7f213c5c8540 <MyVideoStream>)
    at softhddev.c:1965
#20 0x00007f213c3b07e1 in VdpauDisplayHandlerThread () at video.c:9033
#21 0x00007f213c3b1561 in VideoDisplayHandlerThread (dummy=0x0) at video.c:9914
#22 0x00007f213f8360a4 in start_thread () at /lib64/libpthread.so.0
#23 0x00007f213e25f7fd in clone () at /lib64/libc.so.6

Display More

vdr-lan:/video/src # svdrpsend plug

Code

220 vdr-lan SVDRP VideoDiskRecorder 2.0.6; Sat Nov 29 19:03:39 2014; UTF-8
214-Available plugins:
214-softhddevice v0.6.1rc1 - Ein Software und GPU emulieres HD-Gerät
214-radio v1.0.0 - Hintergr.Bilder/RDS-Text für Radiosender
214-fritzbox v1.5.3 - Fritz Plugin für die AVM Fritz!Box
214-osdteletext v0.9.4 - Zeigt den Videotext auf dem OSD an
214-femon v2.0.3 - DVB Signal Informationsanzeige (OSD)
214-extrecmenu v1.2.4 - Erweitertes Aufzeichnungs-Menü
214-streamdev-server v0.6.1-git - VDR Streaming Server
214-graphtftng v0.4.10 - VDR OSD on TFT
214-cdplayer v1.2.2 - CD-Player
214-externalplayer v0.3.3 - launch external players
214-menuorg v0.4.4 - Reorganisiert das Haupmenü

Display More

rofafor · November 29, 2014 at 8:34 PM

Quote from Ulrich Eckhardt

This is my first attempt to set up a VDR with intel drivers so I am not sure if all my versions are correct

Please, disable first VDPAU in the softhddevice plugin as your backtrace seems to be using it instead of VAAPI.

Ulrich Eckhardt · November 30, 2014 at 2:05 PM

Quote from rofafor

Please, disable first VDPAU in the softhddevice plugin as your backtrace seems to be using it instead of VAAPI.

Thanks, I had an old parameter for using VDPAU in my setup. With a single screen setup I get now a picture.

My thermaltake case has a built in LCD-Display which I had already used for graphtft. This display is connected to the VGA output. I have configured this as screen 1 and the TV as screen 0 connected to the HDMI ouput. After this change I got strange effects. Initially after start up i got a completely black picture on the VGA output and a red picture on the TV (HDMI output). After moving the mouse I got a picture on both outputs. So this looks like an intel driver problem to me. Has anybody a working xorg.conf? Attached is my current xorg.conf

Getting a simple dual screen setup with two X11 screens is really annoying with the intel drivers , I had to specify the resolutions to use at both displays instead that the driver chooses the correct resolution as send by the EDID information send by the monitors.

johns · December 2, 2014 at 10:55 AM

Quote from Copperhead

MotionCompensated deinterlacing doesn't seem to work at all. Massive interlace artifacts in fast moving objects.

MotionAdaptive deinterlacing works quite good with first field: 2 and second field 1. But if you take a closer look at moving texts you can spot some juddering.

Does SoftHdDevice says it is supported? IIRC there was no fallback to worser Deinterlacer.

Johns

Space · December 3, 2014 at 12:46 AM

Hello everyone,

I just want to report status. I am running current GIT master of libva + libva-intel-driver on my Celeron G1610 (Ivy Bridge based) and both VAAPI-BOB and VAAPI-MADI are running "smooth". For VAAPI-MADI I have to set first field to 1 and second field to 0.

But there are some artifacts with MADI deinterlacing (can be easily seen in N24) but this seems to be driver based (XBMC people noticed the same problems). MCDI is not advertised with GIT master:

Code

Dec 03 00:40:42 [vdr] video/vaapi: noise reduction supported_
Dec 03 00:40:42 [vdr] video/vaapi: 0,00 - 1,00 ++ 0,03 = 0,50_
Dec 03 00:40:42 [vdr] Enabling denoise filter (pos = 0)_
Dec 03 00:40:42 [vdr] video/vaapi: deinterlacing supported_
Dec 03 00:40:42 [vdr] video/vaapi: bob deinterlace supported_
Dec 03 00:40:42 [vdr] video/vaapi: motion adaptive deinterlace supported_
Dec 03 00:40:42 [vdr] Enabling Deint (pos = 1)_
Dec 03 00:40:42 [vdr] Allocating 1 forward reference surfaces for postprocessing_
Dec 03 00:40:42 [vdr] Allocating 0 backward reference surfaces for postprocessing_

There is a branch (19.vpp.adi) by Gwenole Beauchesne which is supposed to contain some fixes regarding deinterlacing on Sandy / Ivy Bridge and even enable MCDI on Ivy Bridge.

But with that driver version I get no deinterlacing at all at MADI/MCDI setting even though it's reported as supported according to VDR log. Next weekend I should be able to provide those logs again.

So from my side: thanks a lot for your hard work!

Best regards, Space

gentoo-vdr · January 12, 2015 at 7:31 PM

Some more (late) feedback from me:
The stuttering problem after setting and moving some editing marks, which I had with v9, is now gone.
MCDI is now usable and looks very good (I had to reset the FirstField and SecondField setting to 0 for both MADI and MCDI). The ticker on N24 is very good readable. Maybe it could be a little bit smoother but that's complaining on a high level.
All in all I am very satisfied with the current status and use it in my productive environment. Thank you very much Antti!!!
By the way, I have a Celeron G1840 and I am using libva-1.5.0, libva-intel-driver-1.5.0 and xf86-video-intel-2.99.917.

TheChief · January 19, 2015 at 1:30 PM

Vielleicht komm ich ja die Tage auch mal zum Testen, aber vorher noch ein paar Fragen:

1. Ist der Patch noch aktuell oder gibts schon einen neueren?
2. Kann ich libva parallel installieren, ohne das es Probleme gibt, wenn ich später meine NVidia Karte in den Rechner stecke? Oder muss ich die dann wieder deinstallieren.
3. Was kann/muss ich auch den Debian-Repos installieren und was von Hand installieren?

Den intel-Treiber nehme ich an, aus dem git im Ausgangspost.

Was ist mit dem Rest?

Code

root@homeserver:~# apt-cache policy libva1
libva1:
  Installiert:           1.4.1-1
  Installationskandidat: 1.4.1-1
  Versionstabelle:
 *** 1.4.1-1 0
        500 http://mirror.switch.ch/ftp/mirror/debian/ jessie/main amd64 Packages
        100 /var/lib/dpkg/status


root@homeserver:~# apt-cache policy ffmpeg
ffmpeg:
  Installiert:           10:2.5.2-dmo1
  Installationskandidat: 10:2.5.2-dmo1
  Versionstabelle:
 *** 10:2.5.2-dmo1 0
        500 http://www.deb-multimedia.org/ jessie/main amd64 Packages
        100 /var/lib/dpkg/status

Display More

Was ist mit xf86-video-intel 2.99.916? Da finde ich nur xserver-xorg-video-intel in der Version 2:2.21.15-2+b2.

Achso, softhddevice patchen ist klar.

Grüsse
TheChief

mikeM · January 19, 2015 at 2:58 PM

Hallo TheChief,
im MLD Portal ist da gerade eine Diskussion drum im Gange. Dort wurde auch der V10 Patch in die Distrie integriert. (Wäre für Dich auch eine Option, mit der sich leicht testen liesse, ohne alles neu machen zu müssen) Leider hat der scheinbar Nebenwirkungen auf den NVidia Part. Die Karten scheinen durchweg bis zu 20 % der Frames zu verlieren. Dies ist auch in Microrucklern zu sehen. Ist das hier eigentlich bekannt? Oder kann es hierfür auch noch andere Gründe geben? Der NVidia, der dort verwendet wird ist der aktuellste 346 er. Und der Kernel ist ein 3.16 er.
Gruß
Michael

TheChief · January 19, 2015 at 3:07 PM

sofhddevice patchen ist kein Problem. Ich will mir nur nicht zwangsläufig mein laufendes System zerschiessen. Daher wollt ich mich mal vorher kundig machen, was benötigt wird und was das für Konsequenzen auf den Betrieb mit Nvidia hat.

Quote from mikeM

Leider hat der scheinbar Nebenwirkungen auf den NVidia Part. Die Karten scheinen durchweg bis zu 20 % der Frames zu verlieren.

Du meinst, wenn ich softhddevice patche, hat das auch Aufwirkungen auf vdpau? Oder versteh ich da jetzt was falsch?

mikeM · January 19, 2015 at 3:39 PM

So hab ich es gemeint. In der MLD ist seit einiger Zeit auch dieser Patch drin. Seit dem ist es so, das die NVidia bstückten Systeme bis zu 20% Paketverluste haben. Auch die Bezeichnungen für temporal_spatial wurden angepasst, was für einiges an Verwirrung sorgte. Außerdem ist Softhddevice mit dem Patch sehr "gesprächig", was in Verbindung mit den gedropten Frames zu viel Schrott im Log geführt hat.

Gruß
Michael

fnu · January 19, 2015 at 3:51 PM

Quote from mikeM

Auch die Bezeichnungen für temporal_spatial wurden angepasst,

Nein, nicht angepasst, bei VDPAU Betrieb sollte da nachwievor "temporal_spatial" stehen, bei Intel Betrieb eben "Motion Compensated Deinterlacing". Aber das ist eben noch nicht fehlerfrei in SHDD integriert, daher würde ich das eher in einem testing/unstable Branch freigeben. Wer es produktiv nutzen will, kann das ja dennoch schon tun ...

Regards
fnu

mikeM · January 19, 2015 at 3:54 PM

fnu Hast Du denn auch mal Frame Verluste festgestellt mit der gepatchten Version mit NVidia Karten?

Gruß
Michael

fnu · January 19, 2015 at 3:57 PM

Quote from mikeM

fnu Hast Du denn auch mal Frame Verluste festgestellt mit der gepatchten Version mit NVidia Karten?

Ich hatte mal eine gepatchte Version mit Nvidia im Einsatz und keine Probleme festgestellt, GT630. Es ist aber nicht auszuschließen das das passieren kann, daher mein gebetsmühlenartiger Aufruf mitzutesten und Feedback zu liefern, sonst wird das nix.

Statt dessen wird das bei MLD gleich integriert? Wow, verquere Welt ...

Regards
fnu

[Announce] VA-API/VPP Patch for vdr-plugin-softhddevice - v10

Participate now!

Tags