AMD's UVD2-based XvBA Finally Does Something On Linux

[Kano]

Well i tested vlc+gl output with a slower system (E8400+HD4550) then the systemload was really too high. A i7-880+HD5670 does not really need video accelleration but it worked however. opengl video output is something to avoid with vlc, xv (which works with nvidia) is much faster - for dvd size it is fast enough with gl. Basically with a fast cpu the open source driver + xv is less problematic. When you have the direct comparision of nv and ati cards in the same system you know definitely what works better - just that even chip cpus are getting so fast now that gpu accelleration is less important than before.

[krionius]

Seems that 10.10-444 ubuntu driver with xorg 1.9 is working a bit more better for hw accelerated h264 playback on my HD 5650 card, at least now the picture is recognizable even if its still garbled... Hope to see a release that actually works... anyone else withe evergreen and better luck?

Screen shot:

http://ompldr.org/vNW44cA

[gbeauche]

I wrote a test-app, where, while playing a file, every x frames i destroy surface(s) and create new with same effect as above.

Could you please send your test file & source?

It seems, as vaDestroySurface does not really freeing resources.

Yes, the actual XvBA resources are retained until the context they are associated with is destroyed. This is some particular design of XvBA, but it could also be a regression when I cleaned up the code.

[Kano]

@krionius

You have to use vlc. mplayer is not possible.

[tbshl-vdr]

Could you please send your test file & source?

I can't attach files (why?), so you can find it here:

http://www.vdr-portal.de/board/thread.php?postid=943496#post943496

Testfile - i have no special file, it should be reproducable with any format.

Yes, the actual XvBA resources are retained until the context they are associated with is destroyed. This is some particular design of XvBA

Wich context is meant here, especially surfaces in case of software-decoding?
If hw-decode is used, context (from vaCreateContext) is destroyed before creating new surfaces.

In my test-app, with key 'r' you can "restart" playing (line 1467), or uncomment #define at line 1440, wich will simulate this.
When uncommented #define in line 1276, only surfaces (and images) are recreated, no context - i used this for sw-decoded content (or uncommntet line 678 to force sw-decoding).

If i run the app and use 'r' to "restart", after about 15 times, screen remain blank, and a few loops later, vaPutSurface gives me an error.
This is not the case, if i (re)open XDisplay (in init_wnd uncomment line 1537, also in destroy_wnd line 1624, in main, comment line 1654).

Btw., with xvba 0.7.4 and Catalyst 10.9, i can't play files using hw-decoding, must go back to 10.7.

Hope you can do something with this infos, if questions, please tell me.

[DivineGrace]

Basically with a fast cpu the open source driver + xv is less problematic. When you have the direct comparision of nv and ati cards in the same system you know definitely what works better - just that even chip cpus are getting so fast now that gpu accelleration is less important than before.

To a certain extent, I agree with you. However, with mobile computing, XvBA is essential for power conservation. For example, if I can keep my dual-core down to 800mhz while watching 720p content, my notebook will not get as hot and will be much more energy efficient.
With fast cpus on my desktop, I don't care much for power conservation nor do I need gpu-accelerated video; ffmpeg-mt takes care of all that.

[gbeauche]

Wich context is meant here, especially surfaces in case of software-decoding?

Underlying XvBA context. You can't create an XvBA surface without a context. But this is an implementation detail you don't need, neither care, to know about.

If hw-decode is used, context (from vaCreateContext) is destroyed before creating new surfaces.

In my test-app, with key 'r' you can "restart" playing (line 1467), or uncomment #define at line 1440, wich will simulate this.
When uncommented #define in line 1276, only surfaces (and images) are recreated, no context - i used this for sw-decoded content (or uncommntet line 678 to force sw-decoding).

In summary, the files you provided exhibit the problem without uncommenting anything, i.e. as is included in the sources?

Btw., with xvba 0.7.4 and Catalyst 10.9, i can't play files using hw-decoding, must go back to 10.7.

On which GPU?

[gbeauche]

In my test-app, with key 'r' you can "restart" playing (line 1467), or uncomment #define at line 1440, wich will simulate this.

I am sorry but I get a crash instead:

0x08049868 in av_release_buffer (avctx=0x8071430, pic=0x808c870) at va.c:326
326 if(ctxt->surfaces[i].id == sid)
(gdb) bt
#0 0x08049868 in av_release_buffer (avctx=0x8071430, pic=0x808c870) at va.c:326
#1 0x004b5e7f in MPV_common_end () from /usr/lib/i686/cmov/libavcodec.so.52
#2 0xbffff8b8 in ?? ()

[gbeauche]

I am sorry but I get a crash instead:

0x08049868 in av_release_buffer (avctx=0x8071430, pic=0x808c870) at va.c:326
326 if(ctxt->surfaces[i].id == sid)
(gdb) bt
#0 0x08049868 in av_release_buffer (avctx=0x8071430, pic=0x808c870) at va.c:326
#1 0x004b5e7f in MPV_common_end () from /usr/lib/i686/cmov/libavcodec.so.52
#2 0xbffff8b8 in ?? ()

Here is a better trace.

Program received signal SIGSEGV, Segmentation fault.
0x08049868 in av_release_buffer (avctx=0x80713d0, pic=0x808c890) at va.c:326
326 if(ctxt->surfaces[i].id == sid)
(gdb) bt
#0 0x08049868 in av_release_buffer (avctx=0x80713d0, pic=0x808c890) at va.c:326
#1 0x004b5e7f in free_frame_buffer (s=0x80a4ec0) at /home/gb/svn/packages/build-area/ffmpeg-0.6/libavcodec/mpegvideo.c:206
#2 free_picture (s=0x80a4ec0) at /home/gb/svn/packages/build-area/ffmpeg-0.6/libavcodec/mpegvideo.c:331
#3 MPV_common_end (s=0x80a4ec0) at /home/gb/svn/packages/build-area/ffmpeg-0.6/libavcodec/mpegvideo.c:746
#4 0x003d8599 in ff_h264_decode_end (avctx=0x80713d0) at /home/gb/svn/packages/build-area/ffmpeg-0.6/libavcodec/h264.c:3365
#5 0x005785da in avcodec_close (avctx=0x80713d0) at /home/gb/svn/packages/build-area/ffmpeg-0.6/libavcodec/utils.c:710
#6 0x0804a711 in close_avctxt (ctxt=0xbffffa08) at va.c:574
#7 0x0804cdb7 in playfile (fname=0xbffffde6 "/mnt/videos/cavium/dirt2.480p30.2Mbps.264", dpy=0xbffffbd4, vadpy=0xbffffbd0,
drawable=0xbffffbcc, wnd_w=1920, wnd_h=1080) at va.c:1470
#8 0x0804d229 in main (argc=<value optimized out>, argv=0x804d356) at va.c:1658

[gbeauche]

In my test-app, with key 'r' you can "restart" playing (line 1467), or uncomment #define at line 1440, wich will simulate this.
When uncommented #define in line 1276, only surfaces (and images) are recreated, no context - i used this for sw-decoded content (or uncommntet line 678 to force sw-decoding).

BTW, just before the crash, I could notice that, on my side (xvba-video), all surfaces are destroyed:

destroying surface 0x03000000
xvba_video: XVBADestroySurface(): surface 0x9165730
destroying surface 0x03000001
xvba_video: XVBADestroySurface(): surface 0x9165a48
destroying surface 0x03000002
xvba_video: XVBADestroySurface(): surface 0x9165d60
destroying surface 0x03000003
xvba_video: XVBADestroySurface(): surface 0x9166078
xvba_video: XVBADestroySurface(): surface 0x94d69a8