[MPlayer-dev-eng] MPlayer/MEncoder crash on using -vf scale and expand together. Fix attached.

[Ivan Kalvachev]

On 10/29/11, Alex C. <alexc.xander at yahoo.in> wrote:
> On Sat, 2011-10-29 at 17:59 +0300, Ivan Kalvachev wrote:
>> I don't hear any objections, so in few hours I'll commit it with the
>> following message:
>> ------------------------
>> Ensure aligned stride for vf_scale output temp buffer.
>>
>> The MP_IMGFLAG_PREFER_ALIDNED_STRIDE is honored only
>> with MP_IMGTYPE_EXPORT, but vf_scale uses only MP_IMGTYPE_TEMP.
>> As result stride won't be rounded up to a multiple of 16.
>>
>> Use MP_IMGFLAG_ACCEPT_ALIDNED_STRIDE instead, as it works for
>> all memory allocated image types..
>>
>> Discovery and initial patch by alexc.xander.yahoo.in.
>> Workarounds segfault in swscale sse3 code due to unaligned access.
>
> Are you sure that shouldn't be MP_IMGFLAG_ACCEPT_ALIGNED? (D->G) :-)

MPlayer already uses av_malloc that is using memaligned internally.
So the image planes are always allocated aligned.

>
> Seriously,  though, it  would  be better  to  replace libswscale  with
> another, cleaner,  scaling engine, even  if a little  slower. Possible
> alternatives  would include  y4mscale and  AVISynth scaling,  and I've
> also heard good things about DScaler. Maybe we can cherry-pick those.

I suspend this and mine patch.
I think I found mistake in my logic. The current flags should work for
all types that are numerically bigger than EXPORT_, aka are not
exported. So the reason for not aligning the stride may be exactly the
reason you pointed out, using of vo that doesn't accept stride.