[FFmpeg-devel] [PATCH] avformat/yuv4mpegdec: simplify math

[Paul B Mahol]

On 5/2/18, wm4 <nfxjfg at googlemail.com> wrote:
> On Tue,  1 May 2018 17:48:32 +0200
> Paul B Mahol <onemda at gmail.com> wrote:
>
>> This one actually works with hd1080 y4m files when seeking backwards.
>>
>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>> ---
>>  libavformat/yuv4mpegdec.c | 8 +++-----
>>  1 file changed, 3 insertions(+), 5 deletions(-)
>>
>> diff --git a/libavformat/yuv4mpegdec.c b/libavformat/yuv4mpegdec.c
>> index 8662a42a4c..de3d5637fe 100644
>> --- a/libavformat/yuv4mpegdec.c
>> +++ b/libavformat/yuv4mpegdec.c
>> @@ -317,11 +317,9 @@ static int yuv4_read_seek(AVFormatContext *s, int
>> stream_index,
>>      AVStream *st = s->streams[0];
>>      int64_t pos;
>>
>> -    pos = av_rescale_rnd(pts * s->packet_size,
>> -                         st->time_base.num,
>> -                         st->time_base.den * s->packet_size,
>> -                         (flags & AVSEEK_FLAG_BACKWARD) ? AV_ROUND_DOWN :
>> AV_ROUND_UP);
>> -    pos *= s->packet_size;
>> +    if (flags & AVSEEK_FLAG_BACKWARD)
>> +        pts -= 1;
>> +    pos = pts * s->packet_size;
>>
>>      if (avio_seek(s->pb, pos + s->internal->data_offset, SEEK_SET) < 0)
>>          return -1;
>
> That seems questionable too. Though I'm not sure if the previous code
> is correct.
>
> In theory, you're supposed to do this:
> 1. Determine seek target as accurately as possible
> 2. If the seek target is not exact then:
> 2a. if AVSEEK_FLAG_FORWARD is set, then round to the next frame
> 2b. otherwise, round to the previous frame
>
> But if pts is in the stream timebase, then 2. can never happen. (And
> also, I have no idea what the code removed by this patch does.)
>
> Regarding mpv, can you just send me a reproducible test case?

Download any hd1080 derf sample and play it until end and than do backstepping,
with --no-correct-pts and without.