[FFmpeg-cvslog] r19901 - trunk/tests/vsynth.regression.ref

[Baptiste Coudurier]

On 09/17/2009 12:46 PM, Reimar D?ffinger wrote:
> On Thu, Sep 17, 2009 at 09:33:47PM +0200, Vitor Sessak wrote:
>> reimar wrote:
>>> Author: reimar
>>> Date: Thu Sep 17 20:57:22 2009
>>> New Revision: 19901
>>>
>>> Log:
>>> Update dnxhd regression test to match the majority of FATE machines (which
>>> also have a much saner stddev).
>>
>> I don't know about the stddev, but valgrind is not happy about this test
>> (reading uninited memory). This is not a false-positive, since the
>> md5sum differs for different --malloc-fill values.
>
> I am well aware there is an issue, I suspected so upon my initial
> commit.
> However this updated value represents some kind of "goal" due to the
> much better stddev compared to the previous value and thus makes more
> sense.
> Apart from this one reason I added the test is that dnxhd is similar to
> DV in one more aspect: the encoder does not properly initialize memory
> it needs to.
> While unfortunately failing regression test seem not particularly
> motivate to fix things, at least it will be documented that DV and DNxHD
> encoders are not reliable.

Huh, they are reliable, they just don't memset 0 aeras that are useless 
(reserved), and in the real world this cause absolutely no problems.

They just don't produce bitexact output if memory is not initialized.
That's the burden of many fixed frame size codecs.

-- 
Baptiste COUDURIER
Key fingerprint                 8D77134D20CC9220201FC5DB0AC9325C5C1ABAAA
FFmpeg maintainer                                  http://www.ffmpeg.org