[Ffmpeg-devel] [PATCH] MJPEG wrongly displayed withverticalYUV422
Mon Oct 30 23:09:04 CET 2006
> Hi Guillaume,
> On Mon, 2006-10-30 at 22:31 +0100, Guillaume POIRIER wrote:
>> > Regression tests do not work when swscaler is enabled. This is a known
>> > issue (not easily fixable, since fixing it would require having 2
>> > different copies of the .ref files)
>> if we want to track down regression test some day or the other, we
>> need to have a reference hash.
>> IMHO, we need to commit one and let the person who runs the regression
>> tests use the relevant version. Maybe we can refine later on, but
>> surely having no way to check regression isn't a good idea!
>> Now, maybe there are other problems such as: mmx, altivec, etc...
>> implementations have different rounding and therefore different hash,
>> which also isn't gonna make things easy.... well...
> I agree. I tried to find a way to have regression tests working both
> with and without swscaler, but I could only come up with this patch:
> which is not acceptable (I agree with Diego and Michael). If anyone has
> better ideas...
The tests are very sensitive, they require bit exact results, what about
using a more relaxed criteria, something like if the PSNR is good enough
then test is passed.
It looks easy to create a "psnr" vhook to compare the reference video
to the transcoded one.
More information about the ffmpeg-devel