On Mon, 25. Nov 01:50, Michael Niedermayer wrote:
> On Sun, Nov 24, 2019 at 11:29:18AM -0500, Andriy Gelman wrote:
> > On Sun, 24. Nov 00:02, Michael Niedermayer wrote:
> > > On Tue, Oct 15, 2019 at 10:50:39PM -0400, Andriy Gelman wrote:
> > > > From: Andriy Gelman <andriy.gel...@gmail.com>
> > > >
> > > > Fixes #7799
> > > >
> > > > Currently, the mp4toannexb filter always inserts the same extradata at
> > > > the start of the first IRAP unit. As in ticket #7799, this can lead to
> > > > decoding errors if modified parameter sets are signalled in-band.
> > > >
> > > > Decoding errors/visual artifacts are also present in the following
> > > > fate-suite/hevc-conformance datasets for hevc->mp4->hevc conversion:
> > > > -RAP_B_Bossen_1.bit
> > > > -RPS_C_ericsson_5.bit
> > > > -SLIST_A_Sony_4.bit
> > > > -SLIST_B_Sony_8.bit
> > > > -SLIST_C_Sony_3.bit
> > > > -SLIST_D_Sony_9.bit
> > > > -TSKIP_A_MS_2.bit
> > > >
> > > > This commit solves these errors by keeping track of VPS/SPS/PPS
> > > > parameter sets
> > > > during the conversion. The correct combination is inserted at the start
> > > > of the first IRAP. SEIs from extradata are inserted before each IRAP.
> > > >
> > > > This commit also makes an update to the hevc-bsf-mp4toannexb fate test
> > > > since the result before this patch contained duplicate parameter sets
> > > > in-band.
> > > > ---
> > > > libavcodec/hevc_mp4toannexb_bsf.c | 503 ++++++++++++++++++++++++++++--
> > > > tests/fate/hevc.mak | 2 +-
> > > > 2 files changed, 472 insertions(+), 33 deletions(-)
> [...]
> > >
> > > [...]
> > > > +/*
> > > > + * Function converts mp4 access unit into annexb
> > > > + * Output packet structure
> > > > + * VPS, SPS, PPS, [SEI(from extradata)], [SEI_PREFIX(from access
> > > > unit)], IRAP, [SEI_SUFFIX]
> > > > + * or
> > > > + * [SEI_PREFIX (from access unit)], [PPS (if not already signalled)],
> > > > VCL(non-irap), [SEI_SUFFIX]
> > > > + */
> > > > static int hevc_mp4toannexb_filter(AVBSFContext *ctx, AVPacket *out)
> > > > {
> > > > HEVCBSFContext *s = ctx->priv_data;
> > > > AVPacket *in;
> > > > - GetByteContext gb;
> > > > -
> > > > - int got_irap = 0;
> > > > - int i, ret = 0;
> > > > + H2645Packet pkt;
> > > > + int i, j, prev_size, ret;
> > > > + int irap_done;
> > > >
> > > > ret = ff_bsf_get_packet(ctx, &in);
> > > > if (ret < 0)
> > > > return ret;
> > > >
> > > > + /* output the annexb nalu if extradata is not parsed*/
> > > > if (!s->extradata_parsed) {
> > > > av_packet_move_ref(out, in);
> > > > av_packet_free(&in);
> > > > return 0;
> > > > }
> > > >
> > >
> > > > - bytestream2_init(&gb, in->data, in->size);
> > >
> > > Is this really better without using an existing bytestream* API ?
> >
> > If I use the API, then I'd have to call bytestraem2_init for each nal unit.
> > I
> > also don't use the bytestream2_get_byte function. It seemed simpler to use
> > the WRITE_NAL macro.
> >
> > But maybe I've misunderstood your question.
>
> i had nothing really specific in mind, just the feeling that using none of
> the existing APIs there is a bit odd.
>
> but more specifically, what about the write side ?
If I use the bytestream* API, then I'd have to add extra boiler plate code each
time I resize the output with av_grow_packet(). I'd have to add something like:
prev_size = bytestream2_tell_p(...);
bytestream2_init_writer(...);
bytestream2_skip_p(..., prev_size);
Or maybe the API needs an extra function that reinitializes the pointers but
keeps the current state of the writer.
--
Andriy
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
