[FFmpeg-devel] [PATCH][RFC] Interpretation of duration field in AVI super index chunk

[Michael Niedermayer]

On Wed, Feb 25, 2015 at 05:35:33PM +0100, Tobias Rapp wrote:
> On 24.02.2015 16:33, Tobias Rapp wrote:
> >Hi,
> >
> >I am currently trying to interpret the index data of HuffYuv/PCM AVI
> >files written by FFmpeg. If the file is larger than 2GiB an "AVIX" RIFF
> >chunk is added for each 2GiB-block, each of these blocks gets an OpenDML
> >chunk index "ix00" and an OpenDML super index chunk "indx" is written
> >for each stream that contains the file offset of the according chunk
> >indexes together with its size and total duration.
> >
> >For this super index "duration" entry the documentation I have found
> >states that it shall be the "time span in stream ticks" [1]. FFmpeg
> >seems to write the chunk count instead which is correct for video
> >streams where one data chunk "00dc" corresponds to one frame. For audio
> >streams one data chunk "01wb" usually contains multiple samples.
> >
> >I guess for audio streams the "duration" should be multiplied by the
> >samples-per-chunk factor in avienc.c to allow third-party applications
> >to jump to the correct AVIX RIFF segment based on a given timestamp.
> >
> >Any comments?
> 
> To illustrate my issue I have attached a patch that makes the
> OpenDML super index chunk consistent with my interpretation of the
> specs. A test with VirtualDub 1.10.4 shows that it also writes the
> audio duration using sample count instead of packet count.
> 
> Using sample count allows to seek the audio stream to the right RIFF
> segment without reading through all OpenDML standard "ix01" chunks.
> 
> Please comment,
> Tobias

>  avienc.c |    9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 09f8c8250ce5ec1bdad79a1bf280028c9d3af376  0001-libavformat-avienc-Fix-duration-for-audio-stream-Ope.patch
> From 3a8e15598fa3e044f3ef65b5063c633cb4b3afed Mon Sep 17 00:00:00 2001
> From: Tobias Rapp <t.rapp at noa-audio.com>
> Date: Wed, 25 Feb 2015 17:10:13 +0100
> Subject: [PATCH] libavformat/avienc: Fix duration for audio stream OpenDML
>  super index
> 
> Fixes the duration field of the OpenDML super index "indx" chunk to
> contain the number of samples instead of the number of packets for
> (linear/PCM) audio streams.
> 
> This matches the OpenDML V1.02 standard text which states that the
> duration field shall contain "time span in stream ticks".
> ---
>  libavformat/avienc.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/libavformat/avienc.c b/libavformat/avienc.c
> index 53c2fe7..b77b295 100644
> --- a/libavformat/avienc.c
> +++ b/libavformat/avienc.c
> @@ -49,6 +49,7 @@ typedef struct AVIIentry {
>  
>  typedef struct AVIIndex {
>      int64_t     indx_start;
> +    int64_t     audio_strm_offset;
>      int         entry;
>      int         ents_allocated;
>      AVIIentry** cluster;
> @@ -91,6 +92,7 @@ static int64_t avi_start_new_riff(AVFormatContext *s, AVIOContext *pb,
>      avi->riff_id++;
>      for (i = 0; i < s->nb_streams; i++) {
>          AVIStream *avist = s->streams[i]->priv_data;
> +        avist->indexes.audio_strm_offset = avist->audio_strm_length;
>          avist->indexes.entry = 0;
>      }
>  
> @@ -476,6 +478,7 @@ static int avi_write_ix(AVFormatContext *s)
>      for (i = 0; i < s->nb_streams; i++) {
>          AVIStream *avist = s->streams[i]->priv_data;
>          int64_t ix, pos;
> +        int au_byterate, au_ssize, au_scale;
>  
>          avi_stream2fourcc(tag, i, s->streams[i]->codec->codec_type);
>          ix_tag[3] = '0' + i;
> @@ -511,7 +514,11 @@ static int avi_write_ix(AVFormatContext *s)
>          avio_skip(pb, 16 * avi->riff_id);
>          avio_wl64(pb, ix);                    /* qwOffset */
>          avio_wl32(pb, pos - ix);              /* dwSize */
> -        avio_wl32(pb, avist->indexes.entry);  /* dwDuration */
> +        ff_parse_specific_params(s->streams[i], &au_byterate, &au_ssize, &au_scale);
> +        if (au_ssize == 0)
> +            avio_wl32(pb, avist->indexes.entry);  /* dwDuration (packet count) */
> +        else

> +            avio_wl32(pb, (avist->audio_strm_length - avist->indexes.audio_strm_offset) / au_ssize);  /* dwDuration (sample count) */

the rounding here is wrong
the index to be useable needs to have segments duration summable
as is this would lead to significangt errors if many of these
durations would be summed up
something like
avist->audio_strm_length / au_ssize - avist->indexes.audio_strm_offset / au_ssize
might avoid that but ive not tested it

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Many things microsoft did are stupid, but not doing something just because
microsoft did it is even more stupid. If everything ms did were stupid they
would be bankrupt already.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150226/bf387bbb/attachment.asc>