[Libav-user] avcodec_decode_video2 doesn't get pal8 palette and how to get "closest" pix_fmt

[Matthew Einhorn]

On Tue, Aug 30, 2011 at 2:01 PM, Matthew Einhorn <moiein2000 at gmail.com> wrote:
> On Thu, Jul 28, 2011 at 8:15 AM, Matthew Einhorn <moiein2000 at gmail.com> wrote:
>> On Thu, Jul 28, 2011 at 6:02 AM, Stefano Sabatini
>> <stefano.sabatini-lala at poste.it> wrote:
>>> On date Tuesday 2011-07-26 13:58:00 -0400, Matthew Einhorn encoded:
>>>> On Sun, Jul 24, 2011 at 9:41 AM, Stefano Sabatini
>>>> <stefano.sabatini-lala at poste.it> wrote:
>>>> > On date Sunday 2011-07-24 04:51:49 -0400, Matthew Einhorn encoded:
>>> [...]
>>>> Upon spending some time with the debugger I've isolated the problem
>>>> into a very weird corner. But first a bit more about my code (partly
>>>> attached). My code is a dll wrapper to the ffmpeg dlls. Upon one dll
>>>> call you create an object for a video and initialize all the
>>>> format/conversion/codec contexts (open function). Then with further
>>>> dll calls you request the next frame. As said, this works fine now for
>>>> all video formats I tested with except pal8 (rawvideo). With pal8,
>>>> calling avcodec_decode_video2(m_pCodecCtx, m_pFrame, &nFrameFinished,
>>>> m_pAVPacket) copies a palette of zero into m_pFrame->data[1]
>>>> (palette_has_changed is also zero). So this has nothing to do with the
>>>> sws_scale, because sws_scale gets a bad palette. So the question is
>>>> why avcodec_decode_video2 doesn't read the palette. The video file
>>>> isn't bad because of the following.
>>>
>>> So far so good.
>>>
>>>> I was able to fix this if before returning from the open function I
>>>> added one call to avcodec_decode_video2 (and of course before that to
>>>> av_read_frame). That is, if I asked ffmpeg to decode the first frame
>>>> before I returned from the function that initialized frames, context
>>>> etc. the palette was read correctly in the first and subsequent frames
>>>> (palette_has_changed was one). But if I requested the first frame
>>>> after returning from my open frame function, in a separate function,
>>>> the palette isn't read properly.
>>>>
>>>> Now, this smells of something going out of context and closed when my
>>>> open function returns. It cannot be my variables because all of my
>>>> variables are created as class variables beforehand which stay put. I
>>>> also don't use any smart pointers or such. So it must be (I think)
>>>> that one of the av alloc functions clears something if I don't decode
>>>> a frame before returning from the function that called the av alloc
>>>> function. I think it's something with the decoder, possibly a buffer?
>>>>
>>>> My dlls are called from the same thread every time they are called and
>>>> the dll doesn't unload or move between calls. Now ffplay does all its
>>>> work from one central main function with calls to other functions (and
>>>> it's not a dll) so that's why I think ffplay doesn't have an issue
>>>> with it.
>>>>
>>>
>>>> Now, I understand that this might be difficult to debug so I'm mostly
>>>> asking for clues and what to look at. I.e. in all the format/codec
>>>> contexts structs is there some function pointer or member variables
>>>> that are responsible for getting the palettes and will help me track
>>>> down the issue?
>>>
>>> Why the
>>>                av_free_packet(m_pAVPacket);
>>>
>>> in cDecodeFrame()?
>>>
>>> This looks suspicious.
>>>
>>
>> The reason for the av_free_packet call after decoding the frame is
>> that that's how the dranger example did it. But ffplay also does the
>> same thing at line 1773:
>> http://www.ffmpeg.org/doxygen/trunk/ffplay_8c-source.html#l01771
>>
>> When I removed the av_free_packet call it introduced a memory leak
>> into the application and the memory use of the app grew with each call
>> to get next frame and it also didn't fix the pal8 palette issue.
>>
>>>> avcodec_decode_video2 ends up calling some function
>>>> pointer so I couldn't follow through the code to see where it's
>>>> actually read. It could also be that the problem is with the the
>>>> zeranoe dlls in which case this might not be the best place to solve
>>>> it, but I doubt it because it works fine for all the other videos.
>>>>
>>>>
>>>> >> In particular, from what I seemed to have read and seen of ffmpeg, for
>>>> >> pal8 AVFrame data[0] is the data, while data[1] is the palette. When
>>>> >> calling avcodec_decode_video2 on a pal8 video, data[0] is indeed data
>>>> >> (bunch of different values), while data[1] is an array with all
>>>> >> elements zero. Indeed, when I edited data[1] to some random values the
>>>> >> sws_scale output image was not black anymore and you could see the
>>>> >> remnants of my picture.
>>>> >>
>>>> >
>>>> >> So I'm wondering, is the video file broken and that's why the palette
>>>> >> doesn't show up? Or did I miss a flag when initializing codec/format
>>>> >> context etc. so that the palette isn't read?
>>>> >
>>>> > AFAIK you don't need any special hacks for working with palette
>>>> > formats.
>>>> >
>>>> >> 2. I'm looking for a function similar to avcodec_find_best_pix_fmt.
>>>> >> What I want is to pass in a list of formats and the function would
>>>> >> return what's the closest format. For example, say the source format
>>>> >> is pal8 and I pass in as possible destination formats: RGB24 and
>>>> >> GRAY8. Then the function should return GRAY8.
>>>> >> avcodec_find_best_pix_fmt would return in that case RGB24 which "is"
>>>> >> the best format, but in this case would waste 2 extra bytes since pal8
>>>> >> is only 8 bytes depth and gray to start with.
>>>> >>
>>>> >> Does a function like this exist? Would it be easy for me to write such
>>>> >> a function using the ffmpeg API? And if so can I get some pointers?
>>>> >
>>>> > Should be easy to hack the logic of avcodec_find_best_pix_fmt() for
>>>> > implementing an avcodec_find_closest_pix_fmt() or such.
>>>> >
>>>>
>>>> I looked through the code for the above functions and I think as is,
>>>> the avcodec_find_best_pix_fmt function should return the closest pix
>>>> format like I want. I think the only reason it doesn't (I think) is
>>>> because the pal8 format in particular might be set wrongly.
>>>>
>>>
>>>> If you look at the pix_fmt_info array that the
>>>> avcodec_find_best_pix_fmt1 func is referring to, you'll see this
>>>> definition for pal8:
>>>> [PIX_FMT_PAL8] = {
>>>>     .is_alpha = 1,
>>>>     .color_type = FF_COLOR_RGB,
>>>>     .depth = 8,
>>>> },
>>>>
>>>> shouldn't it be .color_type = FF_COLOR_GRAY? Because it's set to
>>>> FF_COLOR_RGB, the avcodec get loss function returns a chroma and
>>>> colorspace loss when converting from pal8 to gray8. That's why RGB24
>>>> gets picked over gray8. But I thought that pal8 is already gray (B/W)
>>>> so there shouldn't be any loss? Admittedly, I don't know too much
>>>> about the pix formats.
>>>
>>> Pal8 works by storing a palette in data[1], which maps an integer in
>>> the range 0-255 to an RGBA 32-bits entry.
>>>
>>> The actual chromatic image features can be guessed only by analyzing
>>> the palette itself, in case you have all the RGBA entries set to a
>>> gray color then it will be gray, but in general a PAL8 image will
>>> contain a colored (R != G != B) data.
>>>
>>
>> I didn't know that. The pal8 videos I saw was gray so I assumed all of
>> them were (although if they were all gray a palette wouldn't really be
>> needed...). So avcodec_find_best_pix_fmt should work fine for me then.
>>
>> Thanks,
>> Matt
>>
>
> FYI this seems to have been fixed in one of the recent gits.
> Thanks,
> Matt
>

Oops, to be clear, I mean the issues where the pal8 palette wasn't
being decoded properly with avcodec_decode_video2 is apparently fixed
now.
Matt