[FFmpeg-devel] [PATCH v4 3/3] aadec: improve seeking in mp3 content

[Karsten Otto]

> Am 11.07.2018 um 20:09 schrieb Michael Niedermayer <michael at niedermayer.cc>:
> 
> Signierter PGP-Teil
> On Sat, Jul 07, 2018 at 07:41:29PM +0200, Karsten Otto wrote:
>> MP3 frames may not be aligned to aa chunk boundaries. When seeking,
>> calculate the expected frame offset in the target chunk. Adjust the
>> timestamp and truncate the next packet accordingly.
>> 
>> This solution works for the majority of tested audio material. For
>> some rare encodings with mp3 padding or embedded id3 tags, it will
>> mispredict the correct offset, and at worst skip an extra frame.
>> ---
>> libavformat/aadec.c | 15 ++++++++++++---
>> 1 file changed, 12 insertions(+), 3 deletions(-)
>> 
>> diff --git a/libavformat/aadec.c b/libavformat/aadec.c
>> index e3c03bc522..2b9e4e526c 100644
>> --- a/libavformat/aadec.c
>> +++ b/libavformat/aadec.c
>> @@ -37,6 +37,7 @@
>> #define TEA_BLOCK_SIZE 8
>> #define CHAPTER_HEADER_SIZE 8
>> #define TIMEPREC 1000
>> +#define MP3_FRAME_SIZE 104
>> 
>> typedef struct AADemuxContext {
>>     AVClass *class;
>> @@ -50,6 +51,7 @@ typedef struct AADemuxContext {
>>     int64_t current_chapter_size;
>>     int64_t content_start;
>>     int64_t content_end;
>> +    int seek_offset;
>> } AADemuxContext;
>> 
>> static int get_second_size(char *codec_name)
>> @@ -230,6 +232,7 @@ static int aa_read_header(AVFormatContext *s)
>>     ff_update_cur_dts(s, st, 0);
>>     avio_seek(pb, start, SEEK_SET);
>>     c->current_chapter_size = 0;
>> +    c->seek_offset = 0;
>> 
>>     return 0;
>> }
>> @@ -295,12 +298,13 @@ static int aa_read_packet(AVFormatContext *s, AVPacket *pkt)
>>     if (c->current_chapter_size <= 0)
>>         c->current_chapter_size = 0;
>> 
>> -    ret = av_new_packet(pkt, written);
>> +    ret = av_new_packet(pkt, written - c->seek_offset);
>>     if (ret < 0)
>>         return ret;
>> -    memcpy(pkt->data, buf, written);
>> +    memcpy(pkt->data, buf + c->seek_offset, written - c->seek_offset);
> 
> what if written < c->seek_offset ? or is this impossible ?
> 
In the normal case this cannot happen, as each chapter is a multiple of the
frame size. They could be equal, which would result in an empty packet.
But I am not quite sure about the worst case I mentioned; to be on the safe
side, I can add an extra check for the last chapter block. Will send a new
patch version.

Cheers, Karsten