| Message ID | 20220704141340.493931-3-philip-dylan.gleonec@savoirfairelinux.com |
|---|---|
| State | New |
| Headers | show |
| Series | [FFmpeg-devel,v2,1/2] avcodec/libopusenc: reload packet loss at encode | expand |
| Context | Check | Description |
|---|---|---|
| yinshiyou/make_loongarch64 | success | Make finished |
| yinshiyou/make_fate_loongarch64 | success | Make fate finished |
| andriy/make_x86 | success | Make finished |
| andriy/make_fate_x86 | success | Make fate finished |
| andriy/make_armv7_RPi4 | success | Make finished |
| andriy/make_fate_armv7_RPi4 | success | Make fate finished |
Jul 4, 2022, 16:13 by philip-dylan.gleonec@savoirfairelinux.com: > Adds FEC/PLC support to libopus. The lost packets are detected as a > discontinuity in the audio stream. When a discontinuity is used, this > patch tries to decode the FEC data. If FEC data is present in the > packet, it is decoded, otherwise audio is re-created through PLC. > > This patch is based on Steinar H. Gunderson contribution, and corrects > the pts computation: all pts are expressed in samples instead of time. > This patch also adds an option "decode_fec" which enables or disables > FEC decoding. This option is disabled by default to keep consistent > behaviour with former versions. > > A number of checks are made to ensure compatibility with different > containers. Indeed, video containers seem to have a pts expressed in ms > while it is expressed in samples for audio containers. It also manages > the cases where pkt->duration is 0, in some RTP streams. This patch > ignores data it can not reconstruct, i.e. packets received twice and > packets with a length that is not a multiple of 2.5ms. > > Signed-off-by: Philip-Dylan Gleonec <philip-dylan.gleonec@savoirfairelinux.com> > Co-developed-by: Steinar H. Gunderson <steinar+ffmpeg@gunderson.no> > --- > libavcodec/libopusdec.c | 105 +++++++++++++++++++++++++++++++++++----- > 1 file changed, 94 insertions(+), 11 deletions(-) > > diff --git a/libavcodec/libopusdec.c b/libavcodec/libopusdec.c > index 316ab0f2a7..f5d0e95fc8 100644 > --- a/libavcodec/libopusdec.c > +++ b/libavcodec/libopusdec.c > @@ -44,10 +44,15 @@ struct libopus_context { > #ifdef OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST > int apply_phase_inv; > #endif > + int decode_fec; > + int64_t expected_next_pts; > }; > > #define OPUS_HEAD_SIZE 19 > > +// Sample rate is constant as libopus always output at 48kHz > +static const AVRational opus_timebase = { 1, 48000 }; > + > static av_cold int libopus_decode_init(AVCodecContext *avc) > { > struct libopus_context *opus = avc->priv_data; > @@ -140,6 +145,8 @@ static av_cold int libopus_decode_init(AVCodecContext *avc) > /* Decoder delay (in samples) at 48kHz */ > avc->delay = avc->internal->skip_samples = opus->pre_skip; > > + opus->expected_next_pts = AV_NOPTS_VALUE; > + > return 0; > } > > @@ -160,25 +167,100 @@ static int libopus_decode(AVCodecContext *avc, AVFrame *frame, > int *got_frame_ptr, AVPacket *pkt) > { > struct libopus_context *opus = avc->priv_data; > - int ret, nb_samples; > + uint8_t *outptr; > + int ret, nb_samples = 0, nb_lost_samples = 0, nb_samples_left; > + > + // If FEC is enabled, calculate number of lost samples > + if (opus->decode_fec && > + opus->expected_next_pts != AV_NOPTS_VALUE && > + pkt->pts != AV_NOPTS_VALUE && > + pkt->pts != opus->expected_next_pts) { > + // Cap at recovering 120 ms of lost audio. > + nb_lost_samples = pkt->pts - opus->expected_next_pts; > + nb_lost_samples = FFMIN(nb_lost_samples, MAX_FRAME_SIZE); > + // pts is expressed in ms for some containers (e.g. mkv) > + // FEC only works for SILK frames (> 10ms) > + // Detect if nb_lost_samples is in ms, and convert in samples if it is > + if (nb_lost_samples > 0) { > + if (avc->pkt_timebase.den != 48000) { > + nb_lost_samples = av_rescale_q(nb_lost_samples, avc->pkt_timebase, opus_timebase); > + } > + // For FEC/PLC, frame_size has to be to have a multiple of 2.5 ms > + if (nb_lost_samples % (5LL * opus_timebase.den / 2000)) { > + nb_lost_samples -= nb_lost_samples % (5LL * opus_timebase.den / 2000); > I counted not two, but three different coding styles used in both patches. Fix it. The lost samples count is very wrong for Ogg Opus files, and in general it's simply incorrect. You *always* need to convert PTS into samples properly via the timebase, and you even hardcode a random timebase that happens to correspond to samples in some conditions. Corrupt ogg files have a duration that's 10x longer than the input, whilst disabling FEC or using our native decoder outputs a correct number of samples. You can use ffmpeg -i test.opus -c:a copy -bsf:a noise=amount=0:dropamount=2 -y test2.opus to corrupt Ogg Opus files and test yourself. I used 100% packet loss percentage for my tests.
diff --git a/libavcodec/libopusdec.c b/libavcodec/libopusdec.c index 316ab0f2a7..f5d0e95fc8 100644 --- a/libavcodec/libopusdec.c +++ b/libavcodec/libopusdec.c @@ -44,10 +44,15 @@ struct libopus_context { #ifdef OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST int apply_phase_inv; #endif + int decode_fec; + int64_t expected_next_pts; }; #define OPUS_HEAD_SIZE 19 +// Sample rate is constant as libopus always output at 48kHz +static const AVRational opus_timebase = { 1, 48000 }; + static av_cold int libopus_decode_init(AVCodecContext *avc) { struct libopus_context *opus = avc->priv_data; @@ -140,6 +145,8 @@ static av_cold int libopus_decode_init(AVCodecContext *avc) /* Decoder delay (in samples) at 48kHz */ avc->delay = avc->internal->skip_samples = opus->pre_skip; + opus->expected_next_pts = AV_NOPTS_VALUE; + return 0; } @@ -160,25 +167,100 @@ static int libopus_decode(AVCodecContext *avc, AVFrame *frame, int *got_frame_ptr, AVPacket *pkt) { struct libopus_context *opus = avc->priv_data; - int ret, nb_samples; + uint8_t *outptr; + int ret, nb_samples = 0, nb_lost_samples = 0, nb_samples_left; + + // If FEC is enabled, calculate number of lost samples + if (opus->decode_fec && + opus->expected_next_pts != AV_NOPTS_VALUE && + pkt->pts != AV_NOPTS_VALUE && + pkt->pts != opus->expected_next_pts) { + // Cap at recovering 120 ms of lost audio. + nb_lost_samples = pkt->pts - opus->expected_next_pts; + nb_lost_samples = FFMIN(nb_lost_samples, MAX_FRAME_SIZE); + // pts is expressed in ms for some containers (e.g. mkv) + // FEC only works for SILK frames (> 10ms) + // Detect if nb_lost_samples is in ms, and convert in samples if it is + if (nb_lost_samples > 0) { + if (avc->pkt_timebase.den != 48000) { + nb_lost_samples = av_rescale_q(nb_lost_samples, avc->pkt_timebase, opus_timebase); + } + // For FEC/PLC, frame_size has to be to have a multiple of 2.5 ms + if (nb_lost_samples % (5LL * opus_timebase.den / 2000)) { + nb_lost_samples -= nb_lost_samples % (5LL * opus_timebase.den / 2000); + } + } + } - frame->nb_samples = MAX_FRAME_SIZE; + frame->nb_samples = MAX_FRAME_SIZE + nb_lost_samples; if ((ret = ff_get_buffer(avc, frame, 0)) < 0) return ret; + outptr = frame->data[0]; + nb_samples_left = frame->nb_samples; + + if (opus->decode_fec && nb_lost_samples > 0) { + // Try to recover the lost samples with FEC data from this one. + // If there's no FEC data, the decoder will do loss concealment instead. + if (avc->sample_fmt == AV_SAMPLE_FMT_S16) + ret = opus_multistream_decode(opus->dec, pkt->data, pkt->size, + (opus_int16 *)outptr, + nb_lost_samples, 1); + else + ret = opus_multistream_decode_float(opus->dec, pkt->data, pkt->size, + (float *)outptr, + nb_lost_samples, 1); + + if (ret < 0) { + if (opus->decode_fec) opus->expected_next_pts = pkt->pts + pkt->duration; + av_log(avc, AV_LOG_ERROR, "Decoding error: %s\n", + opus_strerror(ret)); + return ff_opus_error_to_averror(ret); + } + + av_log(avc, AV_LOG_WARNING, "Recovered %d samples with FEC/PLC\n", + ret); + + outptr += ret * avc->ch_layout.nb_channels * av_get_bytes_per_sample(avc->sample_fmt); + nb_samples_left -= ret; + nb_samples += ret; + if (pkt->pts != AV_NOPTS_VALUE) { + frame->pts = pkt->pts - ret; + } + } + + // Decode the actual, non-lost data. if (avc->sample_fmt == AV_SAMPLE_FMT_S16) - nb_samples = opus_multistream_decode(opus->dec, pkt->data, pkt->size, - (opus_int16 *)frame->data[0], - frame->nb_samples, 0); + ret = opus_multistream_decode(opus->dec, pkt->data, pkt->size, + (opus_int16 *)outptr, + nb_samples_left, 0); else - nb_samples = opus_multistream_decode_float(opus->dec, pkt->data, pkt->size, - (float *)frame->data[0], - frame->nb_samples, 0); + ret = opus_multistream_decode_float(opus->dec, pkt->data, pkt->size, + (float *)outptr, + nb_samples_left, 0); - if (nb_samples < 0) { + if (ret < 0) { + if (opus->decode_fec) opus->expected_next_pts = pkt->pts + pkt->duration; av_log(avc, AV_LOG_ERROR, "Decoding error: %s\n", - opus_strerror(nb_samples)); - return ff_opus_error_to_averror(nb_samples); + opus_strerror(ret)); + return ff_opus_error_to_averror(ret); + } + nb_samples += ret; + + if (opus->decode_fec) + { + // Calculate the next expected pts + if (pkt->pts == AV_NOPTS_VALUE) { + opus->expected_next_pts = AV_NOPTS_VALUE; + } else { + if (pkt->duration) { + opus->expected_next_pts = pkt->pts + pkt->duration; + } else if (avc->pkt_timebase.num) { + opus->expected_next_pts = pkt->pts + av_rescale_q(ret, opus_timebase, avc->pkt_timebase); + } else { + opus->expected_next_pts = pkt->pts + ret; + } + } } #ifndef OPUS_SET_GAIN @@ -219,6 +301,7 @@ static const AVOption libopusdec_options[] = { #ifdef OPUS_SET_PHASE_INVERSION_DISABLED_REQUEST { "apply_phase_inv", "Apply intensity stereo phase inversion", OFFSET(apply_phase_inv), AV_OPT_TYPE_BOOL, { .i64 = 1 }, 0, 1, FLAGS }, #endif + { "decode_fec", "Decode FEC data or use PLC", OFFSET(decode_fec), AV_OPT_TYPE_BOOL, { .i64 = 0 }, 0, 1, FLAGS }, { NULL }, };
Adds FEC/PLC support to libopus. The lost packets are detected as a discontinuity in the audio stream. When a discontinuity is used, this patch tries to decode the FEC data. If FEC data is present in the packet, it is decoded, otherwise audio is re-created through PLC. This patch is based on Steinar H. Gunderson contribution, and corrects the pts computation: all pts are expressed in samples instead of time. This patch also adds an option "decode_fec" which enables or disables FEC decoding. This option is disabled by default to keep consistent behaviour with former versions. A number of checks are made to ensure compatibility with different containers. Indeed, video containers seem to have a pts expressed in ms while it is expressed in samples for audio containers. It also manages the cases where pkt->duration is 0, in some RTP streams. This patch ignores data it can not reconstruct, i.e. packets received twice and packets with a length that is not a multiple of 2.5ms. Signed-off-by: Philip-Dylan Gleonec <philip-dylan.gleonec@savoirfairelinux.com> Co-developed-by: Steinar H. Gunderson <steinar+ffmpeg@gunderson.no> --- libavcodec/libopusdec.c | 105 +++++++++++++++++++++++++++++++++++----- 1 file changed, 94 insertions(+), 11 deletions(-)