[Mesa-dev] [PATCH 5/6] vl: Remove most members of pipe_video_decoder

[Andy Furniss]

Maarten Lankhorst wrote:
> Hey Andy,
>
> On 12/19/2011 01:50 PM, Andy Furniss wrote:
>> Maarten Lankhorst wrote:
>>>
>>> Looks like I made the checking a bit too paranoid, causing it to not always decode last few macroblocks because of fear of reaching end of stream..
>>> Does this work? I removed a few paranoid asserts, and if performance is still bad, I'll alter the patch to swap between multiple decode buffers to fix it. That should bring performance back to old level.
>>
>> This version does fix the macroblocks issue.
>>
>> Perf is still poor.
>>
>> The dvd assert is also still present. It doesn't happen with all dvds - and is not a regression caused by this patch.
>>
>> When playing with -cache and skipping around I get a different assert - attached are backtraces from both types.
> Yeah that version didn't fix performance yet, this one should have the same performance as before. However those backtraces miss the interesting parts, could you build with -O0 instead of -O2 for a complete backtrace?

Perf is back to how it was - just good enough (vdpau is not exactly fast 
compared to anything else), xvmc is better, but has an issue (not new) 
where sometimes it looks jittery like the frame order is wrong (only 
seen on HD with my card on low).

I compiled with -O0 but the backtraces are different - maybe there is 
some randomness.

When getting the skip one I managed to get a couple of hangs rather than 
crashes. One of which I could see was trying some strange resize, eg.
instead of

VO: [vdpau] 720x576 => 1024x576 MPEG2 VDPAU acceleration

I could see something like (from memory)

VO: [vdpau] 19x1893 => 2032x2125 MPEG2 VDPAU acceleration

and it hung in the process of enlarging the window.