[MPlayer-dev-eng] branching off rc4

[compn]

On Sat, 10 Jul 2010 23:05:19 +0200, Reimar Döffinger wrote:
>On Sat, Jul 10, 2010 at 03:53:50PM -0400, compn wrote:
>> On Sat, 10 Jul 2010 16:58:13 +0200, Steinar H. Gunderson wrote:
>> >On Fri, Jun 25, 2010 at 12:07:43PM +0200, Attila Kinali wrote:
>> >> Are there any known regressions or bugs that should be fixed
>> >> before a release?
>> >
>> >I ran my codec test suite (mostly binary) against current trunk, and it's
>> >bombing left and right. Something central has changed, although I cannot say
>> >exactly what, nor whether it's something benign (ie. some roundoff somewhere
>> >changed) or sign of a real bug. (FWIW, the test asks MPlayer to decode a
>> >sample to an yuv4mpeg output and then compares its checksum to a known one,
>> >so it will certainly be sensitive to changes in roundoff.)
>> >
>> >Does anyone have any good ideas that would save me from a long bisect?
>> >The tests are:
>> >
>> >Passing: fraps qt3ivx qtavui qtcvid qtprores qtsvq3 rv40win wmv9dmo wmvdmo wmvvc1dmo
>> >Failing (different output): divxds ffh264 indeo5 msuscls qt261 smartsight vp6 vp7 vsswlt wmv8 zmbv
>> 
>> according to my binaries, it was between r30521 and r31372 for
>> vp6. so at least that narrows it down a bit.
>
>What is the difference exactly? It's unlikely there's any bug in real-world
>use with ffh264, I'm quite convinced someone whould have noticed.

i had similar thoughts. i dont notice any difference visually with vp6.
i'm guessing its an swscale change. also wasnt vo yuv4mpeg changed
recently?

try this one maybe:

New Revision: 31320
Log:
Remove colorspace conversion code from -vo yuv4mpeg, -vf scale should
be able to handle this just as well (or better) including interlaced.
If not, this needs to be fixed there instead of duplicating code.


i'm for bit-exact decoding where possible. i wonder if it matched
previously to reference decoders? would it be hard to get md5sum/yuv
from official decoders and add that to the test suite?


>> btw, why not use -vo md5sum?
>
>The way that does the calculation has changed as well, so anything
>except grayscale pictures would be "broken" when just comparing
>its output.

its funny that two ways to test regressions have been changed.

-compn