On Sun, Jul 24, 2016 at 09:55:21PM +0300, Ivan Uskov wrote:
> Hello All,
>
> I have discovered the following issue:
> Latest builds of ffmpeg crashes into the h264.c when *hardware* qsv-based
> h264 decoder uses.
> The crash does appear inside the
>
> int avpriv_h264_has_num_reorder_frames(AVCodecContext *avctx)
> {
> H264Context *h = avctx->priv_data;
> return h && h->ps.sps ? h->ps.sps->num_reorder_frames : 0;
> }
>
> It is obvious, that casting to H264Context cannot be used for qsv decoder
> because there is QSVH2645Context. Similar issue will appear for CUVID
> decoder case (CuvidContext uses), Android MediaCodec H.264 decoder
> (MediaCodecH264DecContext uses), possible another cases existing.
>
> The caller function is
>
> static int has_decode_delay_been_guessed(AVStream *st)
> {
> if (st->codecpar->codec_id != AV_CODEC_ID_H264) return 1;
> if (!st->info) // if we have left find_stream_info then nb_decoded_frames
> won't increase anymore for stream copy
> return 1;
> #if CONFIG_H264_DECODER
> if (st->internal->avctx->has_b_frames &&
> avpriv_h264_has_num_reorder_frames(st->internal->avctx) ==
> st->internal->avctx->has_b_frames)
> return 1;
> #endif
> if (st->internal->avctx->has_b_frames<3)
> return st->nb_decoded_frames >= 7;
> else if (st->internal->avctx->has_b_frames<4)
> return st->nb_decoded_frames >= 18;
> else
> return st->nb_decoded_frames >= 20;
> }
> ...which called by update_initial_timestamps()
>
> Have anybody the idea how it can be correctly fixed?
> Looks like has_decode_delay_been_guessed() should be corrected.this code is not new, this is in git since 4 years commit bafa1c7f383d6c1c0f3d207395fe6a574092b7ac Date: Mon Jul 2 23:16:59 2012 +020 why does it cause a problem now ? why does libavformat use hw h264 decoding ? this seems like a bad idea to me, as libavcodec would use hw decoding too potentially resulting in multiple hw decoders running at the same time for no good reason. the one in libavformat is just used to parse a few header values ideally that should be done in the AVParser for h264 ... [...] -- Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB Why not whip the teacher when the pupil misbehaves? -- Diogenes of Sinope
signature.asc
Description: Digital signature
_______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
