[FFmpeg-devel] [FFmpeg-cvslog] ffplay: only use hardware accelerated SDL texture formats

[Marton Balint]

On Wed, 18 Apr 2018, Michael Niedermayer wrote:

> On Tue, Apr 17, 2018 at 09:50:15PM +0200, Marton Balint wrote:
>>
>>
>> On Tue, 17 Apr 2018, Michael Niedermayer wrote:
>>
>>> On Tue, Apr 17, 2018 at 01:27:48AM +0200, Marton Balint wrote:
>>>>
>>>>
>>>> On Mon, 16 Apr 2018, Michael Niedermayer wrote:
>>>>
>>>>> Hi
>>>>>
>>>>> On Sat, Nov 04, 2017 at 06:28:58PM +0000, Marton Balint wrote:
>>>>>> ffmpeg | branch: master | Marton Balint <cus at passwd.hu> | Sat Oct 28 22:46:08 2017 +0200| [415038f2bd321a3b41564d4e0c6c17d7a096c397] | committer: Marton Balint
>>>>>>
>>>>>> ffplay: only use hardware accelerated SDL texture formats
>>>>>
>>>>> This breaks ffplay playing some files like:
>>>>> ./ffplay fate-suite//cvid/catfight-cvid-pal8-partial.mov -noframedrop
>>>>>
>>>>> The output is completely black since this commit
>>>>
>>>> Seems like a bug in swscale (pal8 -> bgra conversion), the alpha is 0
>>>> instead of 255.
>>>
>>> the file seems to store alpha = 0
>>> SDL seems to treat alpha=0 different between PAL8 and RGBA, or maybe iam
>>> missing something
>>
>> Old code negotiated RGB24 format for render which silently dropped the alpha
>> channel. This seems to be because of a bug in pick_format:
>>
>> libavfilter/avfiltergraph.c, pick_format():
>> int has_alpha= av_pix_fmt_desc_get(ref->format)->nb_components % 2 == 0;
>>
>> libavutil/pixdesc.c:
>> #define pixdesc_has_alpha(pixdesc) \
>>     ((pixdesc)->nb_components == 2 || (pixdesc)->nb_components == 4 || (pixdesc)->flags & AV_PIX_FMT_FLAG_PAL)
>>
>>
>>>
>>> try this:
>>>
>>> diff --git a/libavformat/qtpalette.c b/libavformat/qtpalette.c
>>> index 666c6b7351..51a134d30a 100644
>>> --- a/libavformat/qtpalette.c
>>> +++ b/libavformat/qtpalette.c
>>> @@ -104,6 +104,7 @@ int ff_get_qtpalette(int codec_id, AVIOContext *pb, uint32_t *palette)
>>>                    avio_r8(pb);
>>>                    b = avio_r8(pb);
>>>                    avio_r8(pb);
>>> +                    a = 0xFF;
>>>                    palette[i] = (a << 24 ) | (r << 16) | (g << 8) | (b);
>>>                }
>>>            }
>>>
>>
>> Obviously this fixes it. Why do we think that the first value is alpha?
>> Parsing it was added in 0d59ae32c23 but according to the QuickTime specs the
>> first int16 should be simply 0, not alpha:
>
> the question here is probably are there files which use this for alpha ?
> If theres no flag or othe to detect this then we could check for the
> spec variant of all being 0 and treat this as all 0xFF alpha
>
> I dont know or remember why this was using it as alpha

I think it was only because the comment in the code said it is alpha when 
it was ignored, and when alpha support was added it made sense to make use 
of it. Maybe Carl remembers, he was the patch author.

I checked the samples we have, and none of the .mov files with palette are 
using a non-zero "alpha", so I'd say we should simply ignore it.

Regards,
Marton