diff mbox series

[FFmpeg-devel,2/2] libfdk-aacdec: Flush delayed samples at the end

Message ID 20220119203837.9047-2-martin@martin.st
State Accepted
Commit c69b1a12bb6231bd648cf29c1d99282e2d5e68d0
Headers show
Series [FFmpeg-devel,1/2] libfdk-aacdec: Add an option for setting the decoder's DRC album mode | expand

Checks

Context Check Description
andriy/make_x86 success Make finished
andriy/make_fate_x86 success Make fate finished
andriy/make_ppc success Make finished
andriy/make_fate_ppc success Make fate finished
andriy/make_aarch64_jetson success Make finished
andriy/make_fate_aarch64_jetson success Make fate finished
andriy/make_armv7_RPi4 success Make finished
andriy/make_fate_armv7_RPi4 success Make fate finished

Commit Message

Martin Storsjö Jan. 19, 2022, 8:38 p.m. UTC 
Also trim off delay samples at the start instead of adjusting pts
to compensate for them; this avoids unwanted offsets if working
with raw samples without considering their pts.
---
 libavcodec/libfdk-aacdec.c | 80 +++++++++++++++++++++++++++++++-------
 1 file changed, 65 insertions(+), 15 deletions(-)

Comments

Andreas Rheinhardt Jan. 21, 2022, 1 p.m. UTC | #1 
Martin Storsjö:
> Also trim off delay samples at the start instead of adjusting pts
> to compensate for them; this avoids unwanted offsets if working
> with raw samples without considering their pts.
> ---
>  libavcodec/libfdk-aacdec.c | 80 +++++++++++++++++++++++++++++++-------
>  1 file changed, 65 insertions(+), 15 deletions(-)
> 
> diff --git a/libavcodec/libfdk-aacdec.c b/libavcodec/libfdk-aacdec.c
> index 93b52023b0..d560e313ca 100644
> --- a/libavcodec/libfdk-aacdec.c
> +++ b/libavcodec/libfdk-aacdec.c
> @@ -58,7 +58,11 @@ typedef struct FDKAACDecContext {
>      int drc_cut;
>      int album_mode;
>      int level_limit;
> -    int output_delay;
> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
> +    int output_delay_set;
> +    int flush_samples;
> +    int delay_samples;
> +#endif
>  } FDKAACDecContext;
>  
>  
> @@ -123,7 +127,12 @@ static int get_stream_info(AVCodecContext *avctx)
>      avctx->sample_rate = info->sampleRate;
>      avctx->frame_size  = info->frameSize;
>  #if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
> -    s->output_delay    = info->outputDelay;
> +    if (!s->output_delay_set && info->outputDelay) {
> +        // Set this only once.
> +        s->flush_samples    = info->outputDelay;
> +        s->delay_samples    = info->outputDelay;
> +        s->output_delay_set = 1;
> +    }
>  #endif
>  
>      for (i = 0; i < info->numChannels; i++) {
> @@ -367,14 +376,31 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
>      int ret;
>      AAC_DECODER_ERROR err;
>      UINT valid = avpkt->size;
> +    UINT flags = 0;
> +    int input_offset = 0;
>  
> -    err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
> -    if (err != AAC_DEC_OK) {
> -        av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
> -        return AVERROR_INVALIDDATA;
> +    if (avpkt->size) {
> +        err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
> +        if (err != AAC_DEC_OK) {
> +            av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
> +            return AVERROR_INVALIDDATA;
> +        }
> +    } else {
> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
> +        /* Handle decoder draining */
> +        if (s->flush_samples > 0) {
> +            flags |= AACDEC_FLUSH;
> +        } else {
> +            return AVERROR_EOF;
> +        }
> +#else
> +        return AVERROR_EOF;
> +#endif
>      }
>  
> -    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer, s->decoder_buffer_size / sizeof(INT_PCM), 0);
> +    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer,
> +                                 s->decoder_buffer_size / sizeof(INT_PCM),
> +                                 flags);
>      if (err == AAC_DEC_NOT_ENOUGH_BITS) {
>          ret = avpkt->size - valid;
>          goto end;
> @@ -390,16 +416,36 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
>          goto end;
>      frame->nb_samples = avctx->frame_size;
>  
> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
> +    if (flags & AACDEC_FLUSH) {
> +        // Only return the right amount of samples at the end; if calling the
> +        // decoder with AACDEC_FLUSH, it will keep returning frames indefinitely.
> +        frame->nb_samples = FFMIN(s->flush_samples, frame->nb_samples);
> +        av_log(s, AV_LOG_DEBUG, "Returning %d/%d delayed samples.\n",
> +                                frame->nb_samples, s->flush_samples);
> +        s->flush_samples -= frame->nb_samples;
> +    } else {
> +        // Trim off samples from the start to compensate for extra decoder
> +        // delay. We could also just adjust the pts, but this avoids
> +        // including the extra samples in the output altogether.
> +        if (s->delay_samples) {
> +            int drop_samples = FFMIN(s->delay_samples, frame->nb_samples);
> +            av_log(s, AV_LOG_DEBUG, "Dropping %d/%d delayed samples.\n",
> +                                    drop_samples, s->delay_samples);
> +            s->delay_samples  -= drop_samples;
> +            frame->nb_samples -= drop_samples;
> +            input_offset = drop_samples * avctx->channels;
> +            if (frame->nb_samples <= 0)
> +                return 0;
> +        }
> +    }
> +#endif
> +
>      if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
>          goto end;
>  
> -    if (frame->pts != AV_NOPTS_VALUE)
> -        frame->pts -= av_rescale_q(s->output_delay,
> -                                   (AVRational){1, avctx->sample_rate},
> -                                   avctx->time_base);
> -
> -    memcpy(frame->extended_data[0], s->decoder_buffer,
> -           avctx->channels * avctx->frame_size *
> +    memcpy(frame->extended_data[0], s->decoder_buffer + input_offset,
> +           avctx->channels * frame->nb_samples *
>             av_get_bytes_per_sample(avctx->sample_fmt));
>  
>      *got_frame_ptr = 1;
> @@ -432,7 +478,11 @@ const AVCodec ff_libfdk_aac_decoder = {
>      .decode         = fdk_aac_decode_frame,
>      .close          = fdk_aac_decode_close,
>      .flush          = fdk_aac_decode_flush,
> -    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
> +    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF
> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
> +                      | AV_CODEC_CAP_DELAY
> +#endif
> +    ,
>      .priv_class     = &fdk_aac_dec_class,
>      .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
>                        FF_CODEC_CAP_INIT_CLEANUP,
> 

When I use the libfdk-aac decoder I get the exact number of samples like
with the native aac decoder (namely number of frames * 1024, as
expected). What makes you believe this is necessary?

- Andreas
Martin Storsjö Jan. 21, 2022, 1:58 p.m. UTC | #2 
On Fri, 21 Jan 2022, Andreas Rheinhardt wrote:

> Martin Storsjö:
>> Also trim off delay samples at the start instead of adjusting pts
>> to compensate for them; this avoids unwanted offsets if working
>> with raw samples without considering their pts.
>> ---
>>  libavcodec/libfdk-aacdec.c | 80 +++++++++++++++++++++++++++++++-------
>>  1 file changed, 65 insertions(+), 15 deletions(-)
>> 
>> diff --git a/libavcodec/libfdk-aacdec.c b/libavcodec/libfdk-aacdec.c
>> index 93b52023b0..d560e313ca 100644
>> --- a/libavcodec/libfdk-aacdec.c
>> +++ b/libavcodec/libfdk-aacdec.c
>> @@ -58,7 +58,11 @@ typedef struct FDKAACDecContext {
>>      int drc_cut;
>>      int album_mode;
>>      int level_limit;
>> -    int output_delay;
>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>> +    int output_delay_set;
>> +    int flush_samples;
>> +    int delay_samples;
>> +#endif
>>  } FDKAACDecContext;
>>
>> 
>> @@ -123,7 +127,12 @@ static int get_stream_info(AVCodecContext *avctx)
>>      avctx->sample_rate = info->sampleRate;
>>      avctx->frame_size  = info->frameSize;
>>  #if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>> -    s->output_delay    = info->outputDelay;
>> +    if (!s->output_delay_set && info->outputDelay) {
>> +        // Set this only once.
>> +        s->flush_samples    = info->outputDelay;
>> +        s->delay_samples    = info->outputDelay;
>> +        s->output_delay_set = 1;
>> +    }
>>  #endif
>>
>>      for (i = 0; i < info->numChannels; i++) {
>> @@ -367,14 +376,31 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
>>      int ret;
>>      AAC_DECODER_ERROR err;
>>      UINT valid = avpkt->size;
>> +    UINT flags = 0;
>> +    int input_offset = 0;
>> 
>> -    err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
>> -    if (err != AAC_DEC_OK) {
>> -        av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
>> -        return AVERROR_INVALIDDATA;
>> +    if (avpkt->size) {
>> +        err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
>> +        if (err != AAC_DEC_OK) {
>> +            av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
>> +            return AVERROR_INVALIDDATA;
>> +        }
>> +    } else {
>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>> +        /* Handle decoder draining */
>> +        if (s->flush_samples > 0) {
>> +            flags |= AACDEC_FLUSH;
>> +        } else {
>> +            return AVERROR_EOF;
>> +        }
>> +#else
>> +        return AVERROR_EOF;
>> +#endif
>>      }
>> 
>> -    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer, s->decoder_buffer_size / sizeof(INT_PCM), 0);
>> +    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer,
>> +                                 s->decoder_buffer_size / sizeof(INT_PCM),
>> +                                 flags);
>>      if (err == AAC_DEC_NOT_ENOUGH_BITS) {
>>          ret = avpkt->size - valid;
>>          goto end;
>> @@ -390,16 +416,36 @@ static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
>>          goto end;
>>      frame->nb_samples = avctx->frame_size;
>> 
>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>> +    if (flags & AACDEC_FLUSH) {
>> +        // Only return the right amount of samples at the end; if calling the
>> +        // decoder with AACDEC_FLUSH, it will keep returning frames indefinitely.
>> +        frame->nb_samples = FFMIN(s->flush_samples, frame->nb_samples);
>> +        av_log(s, AV_LOG_DEBUG, "Returning %d/%d delayed samples.\n",
>> +                                frame->nb_samples, s->flush_samples);
>> +        s->flush_samples -= frame->nb_samples;
>> +    } else {
>> +        // Trim off samples from the start to compensate for extra decoder
>> +        // delay. We could also just adjust the pts, but this avoids
>> +        // including the extra samples in the output altogether.
>> +        if (s->delay_samples) {
>> +            int drop_samples = FFMIN(s->delay_samples, frame->nb_samples);
>> +            av_log(s, AV_LOG_DEBUG, "Dropping %d/%d delayed samples.\n",
>> +                                    drop_samples, s->delay_samples);
>> +            s->delay_samples  -= drop_samples;
>> +            frame->nb_samples -= drop_samples;
>> +            input_offset = drop_samples * avctx->channels;
>> +            if (frame->nb_samples <= 0)
>> +                return 0;
>> +        }
>> +    }
>> +#endif
>> +
>>      if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
>>          goto end;
>> 
>> -    if (frame->pts != AV_NOPTS_VALUE)
>> -        frame->pts -= av_rescale_q(s->output_delay,
>> -                                   (AVRational){1, avctx->sample_rate},
>> -                                   avctx->time_base);
>> -
>> -    memcpy(frame->extended_data[0], s->decoder_buffer,
>> -           avctx->channels * avctx->frame_size *
>> +    memcpy(frame->extended_data[0], s->decoder_buffer + input_offset,
>> +           avctx->channels * frame->nb_samples *
>>             av_get_bytes_per_sample(avctx->sample_fmt));
>>
>>      *got_frame_ptr = 1;
>> @@ -432,7 +478,11 @@ const AVCodec ff_libfdk_aac_decoder = {
>>      .decode         = fdk_aac_decode_frame,
>>      .close          = fdk_aac_decode_close,
>>      .flush          = fdk_aac_decode_flush,
>> -    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
>> +    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF
>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>> +                      | AV_CODEC_CAP_DELAY
>> +#endif
>> +    ,
>>      .priv_class     = &fdk_aac_dec_class,
>>      .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
>>                        FF_CODEC_CAP_INIT_CLEANUP,
>> 
>
> When I use the libfdk-aac decoder I get the exact number of samples like
> with the native aac decoder (namely number of frames * 1024, as
> expected). What makes you believe this is necessary?

The fdk-aac decoder can have, depending on combination of options, some 
amount of extra internal delay, that the libavcodec internal aac decoder 
doesn't have. (It's also possible to set the options in a state where the 
fdk-aac decoder doesn't induce any extra delay.)

Currently, we compensate for that extra delay by just offsetting pts 
backwards, so for a stream with N packets, we return samples with 
timestamps [-delay,N*framesize-delay].

In order not to lose data at the end, we must make the decoder flushable 
and flush up to (delay) samples at the end. And since one doesn't normally 
expect extra delay samples at the start of an AAC decoder output, we also 
trim out the same amount of samples at the start (to simplify for users 
that don't observe the pts, who otherwise are surprised by the stream 
starting from pts -delay instead of at pts 0).

// Martin
Andreas Rheinhardt Jan. 21, 2022, 2:42 p.m. UTC | #3 
Martin Storsjö:
> On Fri, 21 Jan 2022, Andreas Rheinhardt wrote:
> 
>> Martin Storsjö:
>>> Also trim off delay samples at the start instead of adjusting pts
>>> to compensate for them; this avoids unwanted offsets if working
>>> with raw samples without considering their pts.
>>> ---
>>>  libavcodec/libfdk-aacdec.c | 80 +++++++++++++++++++++++++++++++-------
>>>  1 file changed, 65 insertions(+), 15 deletions(-)
>>>
>>> diff --git a/libavcodec/libfdk-aacdec.c b/libavcodec/libfdk-aacdec.c
>>> index 93b52023b0..d560e313ca 100644
>>> --- a/libavcodec/libfdk-aacdec.c
>>> +++ b/libavcodec/libfdk-aacdec.c
>>> @@ -58,7 +58,11 @@ typedef struct FDKAACDecContext {
>>>      int drc_cut;
>>>      int album_mode;
>>>      int level_limit;
>>> -    int output_delay;
>>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>>> +    int output_delay_set;
>>> +    int flush_samples;
>>> +    int delay_samples;
>>> +#endif
>>>  } FDKAACDecContext;
>>>
>>>
>>> @@ -123,7 +127,12 @@ static int get_stream_info(AVCodecContext *avctx)
>>>      avctx->sample_rate = info->sampleRate;
>>>      avctx->frame_size  = info->frameSize;
>>>  #if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>>> -    s->output_delay    = info->outputDelay;
>>> +    if (!s->output_delay_set && info->outputDelay) {
>>> +        // Set this only once.
>>> +        s->flush_samples    = info->outputDelay;
>>> +        s->delay_samples    = info->outputDelay;
>>> +        s->output_delay_set = 1;
>>> +    }
>>>  #endif
>>>
>>>      for (i = 0; i < info->numChannels; i++) {
>>> @@ -367,14 +376,31 @@ static int fdk_aac_decode_frame(AVCodecContext
>>> *avctx, void *data,
>>>      int ret;
>>>      AAC_DECODER_ERROR err;
>>>      UINT valid = avpkt->size;
>>> +    UINT flags = 0;
>>> +    int input_offset = 0;
>>>
>>> -    err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size,
>>> &valid);
>>> -    if (err != AAC_DEC_OK) {
>>> -        av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed:
>>> %x\n", err);
>>> -        return AVERROR_INVALIDDATA;
>>> +    if (avpkt->size) {
>>> +        err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size,
>>> &valid);
>>> +        if (err != AAC_DEC_OK) {
>>> +            av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed:
>>> %x\n", err);
>>> +            return AVERROR_INVALIDDATA;
>>> +        }
>>> +    } else {
>>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>>> +        /* Handle decoder draining */
>>> +        if (s->flush_samples > 0) {
>>> +            flags |= AACDEC_FLUSH;
>>> +        } else {
>>> +            return AVERROR_EOF;
>>> +        }
>>> +#else
>>> +        return AVERROR_EOF;
>>> +#endif
>>>      }
>>>
>>> -    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *)
>>> s->decoder_buffer, s->decoder_buffer_size / sizeof(INT_PCM), 0);
>>> +    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *)
>>> s->decoder_buffer,
>>> +                                 s->decoder_buffer_size /
>>> sizeof(INT_PCM),
>>> +                                 flags);
>>>      if (err == AAC_DEC_NOT_ENOUGH_BITS) {
>>>          ret = avpkt->size - valid;
>>>          goto end;
>>> @@ -390,16 +416,36 @@ static int fdk_aac_decode_frame(AVCodecContext
>>> *avctx, void *data,
>>>          goto end;
>>>      frame->nb_samples = avctx->frame_size;
>>>
>>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>>> +    if (flags & AACDEC_FLUSH) {
>>> +        // Only return the right amount of samples at the end; if
>>> calling the
>>> +        // decoder with AACDEC_FLUSH, it will keep returning frames
>>> indefinitely.
>>> +        frame->nb_samples = FFMIN(s->flush_samples, frame->nb_samples);
>>> +        av_log(s, AV_LOG_DEBUG, "Returning %d/%d delayed samples.\n",
>>> +                                frame->nb_samples, s->flush_samples);
>>> +        s->flush_samples -= frame->nb_samples;
>>> +    } else {
>>> +        // Trim off samples from the start to compensate for extra
>>> decoder
>>> +        // delay. We could also just adjust the pts, but this avoids
>>> +        // including the extra samples in the output altogether.
>>> +        if (s->delay_samples) {
>>> +            int drop_samples = FFMIN(s->delay_samples,
>>> frame->nb_samples);
>>> +            av_log(s, AV_LOG_DEBUG, "Dropping %d/%d delayed
>>> samples.\n",
>>> +                                    drop_samples, s->delay_samples);
>>> +            s->delay_samples  -= drop_samples;
>>> +            frame->nb_samples -= drop_samples;
>>> +            input_offset = drop_samples * avctx->channels;
>>> +            if (frame->nb_samples <= 0)
>>> +                return 0;
>>> +        }
>>> +    }
>>> +#endif
>>> +
>>>      if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
>>>          goto end;
>>>
>>> -    if (frame->pts != AV_NOPTS_VALUE)
>>> -        frame->pts -= av_rescale_q(s->output_delay,
>>> -                                   (AVRational){1, avctx->sample_rate},
>>> -                                   avctx->time_base);
>>> -
>>> -    memcpy(frame->extended_data[0], s->decoder_buffer,
>>> -           avctx->channels * avctx->frame_size *
>>> +    memcpy(frame->extended_data[0], s->decoder_buffer + input_offset,
>>> +           avctx->channels * frame->nb_samples *
>>>             av_get_bytes_per_sample(avctx->sample_fmt));
>>>
>>>      *got_frame_ptr = 1;
>>> @@ -432,7 +478,11 @@ const AVCodec ff_libfdk_aac_decoder = {
>>>      .decode         = fdk_aac_decode_frame,
>>>      .close          = fdk_aac_decode_close,
>>>      .flush          = fdk_aac_decode_flush,
>>> -    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
>>> +    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF
>>> +#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
>>> +                      | AV_CODEC_CAP_DELAY
>>> +#endif
>>> +    ,
>>>      .priv_class     = &fdk_aac_dec_class,
>>>      .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
>>>                        FF_CODEC_CAP_INIT_CLEANUP,
>>>
>>
>> When I use the libfdk-aac decoder I get the exact number of samples like
>> with the native aac decoder (namely number of frames * 1024, as
>> expected). What makes you believe this is necessary?
> 
> The fdk-aac decoder can have, depending on combination of options, some
> amount of extra internal delay, that the libavcodec internal aac decoder
> doesn't have. (It's also possible to set the options in a state where
> the fdk-aac decoder doesn't induce any extra delay.)
> 
> Currently, we compensate for that extra delay by just offsetting pts
> backwards, so for a stream with N packets, we return samples with
> timestamps [-delay,N*framesize-delay].
> 
> In order not to lose data at the end, we must make the decoder flushable
> and flush up to (delay) samples at the end. And since one doesn't
> normally expect extra delay samples at the start of an AAC decoder
> output, we also trim out the same amount of samples at the start (to
> simplify for users that don't observe the pts, who otherwise are
> surprised by the stream starting from pts -delay instead of at pts 0).
> 

Interesting: There is indeed a delay at the start (720 samples in a
quick test) compared to the native AAC decoder.
Furthermore, the current code is buggy, as it believes that
avcodec->time_base to be the time_base of the returned AVFrames (it is
in reality avcodec->pkt_timebase; just test with AAC-in-Matroska for this).
I haven't tested your patches, but I have now realized that there is
indeed an issue. And your patch should also fix the wrong timebase issue.

- Andreas
Martin Storsjö Feb. 3, 2022, 11:53 a.m. UTC | #4 
On Fri, 21 Jan 2022, Andreas Rheinhardt wrote:

> Interesting: There is indeed a delay at the start (720 samples in a
> quick test) compared to the native AAC decoder.
> Furthermore, the current code is buggy, as it believes that
> avcodec->time_base to be the time_base of the returned AVFrames (it is
> in reality avcodec->pkt_timebase; just test with AAC-in-Matroska for this).
> I haven't tested your patches, but I have now realized that there is
> indeed an issue. And your patch should also fix the wrong timebase issue.

Based on discussion on irc with Andreas and James, I think the conclusion 
was that this patch should be fine, so I'll go ahead and push it soon if 
there's no further comments on it.

// Martin
diff mbox series

Patch

diff --git a/libavcodec/libfdk-aacdec.c b/libavcodec/libfdk-aacdec.c
index 93b52023b0..d560e313ca 100644
--- a/libavcodec/libfdk-aacdec.c
+++ b/libavcodec/libfdk-aacdec.c
@@ -58,7 +58,11 @@  typedef struct FDKAACDecContext {
     int drc_cut;
     int album_mode;
     int level_limit;
-    int output_delay;
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+    int output_delay_set;
+    int flush_samples;
+    int delay_samples;
+#endif
 } FDKAACDecContext;
 
 
@@ -123,7 +127,12 @@  static int get_stream_info(AVCodecContext *avctx)
     avctx->sample_rate = info->sampleRate;
     avctx->frame_size  = info->frameSize;
 #if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
-    s->output_delay    = info->outputDelay;
+    if (!s->output_delay_set && info->outputDelay) {
+        // Set this only once.
+        s->flush_samples    = info->outputDelay;
+        s->delay_samples    = info->outputDelay;
+        s->output_delay_set = 1;
+    }
 #endif
 
     for (i = 0; i < info->numChannels; i++) {
@@ -367,14 +376,31 @@  static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
     int ret;
     AAC_DECODER_ERROR err;
     UINT valid = avpkt->size;
+    UINT flags = 0;
+    int input_offset = 0;
 
-    err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
-    if (err != AAC_DEC_OK) {
-        av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
-        return AVERROR_INVALIDDATA;
+    if (avpkt->size) {
+        err = aacDecoder_Fill(s->handle, &avpkt->data, &avpkt->size, &valid);
+        if (err != AAC_DEC_OK) {
+            av_log(avctx, AV_LOG_ERROR, "aacDecoder_Fill() failed: %x\n", err);
+            return AVERROR_INVALIDDATA;
+        }
+    } else {
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+        /* Handle decoder draining */
+        if (s->flush_samples > 0) {
+            flags |= AACDEC_FLUSH;
+        } else {
+            return AVERROR_EOF;
+        }
+#else
+        return AVERROR_EOF;
+#endif
     }
 
-    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer, s->decoder_buffer_size / sizeof(INT_PCM), 0);
+    err = aacDecoder_DecodeFrame(s->handle, (INT_PCM *) s->decoder_buffer,
+                                 s->decoder_buffer_size / sizeof(INT_PCM),
+                                 flags);
     if (err == AAC_DEC_NOT_ENOUGH_BITS) {
         ret = avpkt->size - valid;
         goto end;
@@ -390,16 +416,36 @@  static int fdk_aac_decode_frame(AVCodecContext *avctx, void *data,
         goto end;
     frame->nb_samples = avctx->frame_size;
 
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+    if (flags & AACDEC_FLUSH) {
+        // Only return the right amount of samples at the end; if calling the
+        // decoder with AACDEC_FLUSH, it will keep returning frames indefinitely.
+        frame->nb_samples = FFMIN(s->flush_samples, frame->nb_samples);
+        av_log(s, AV_LOG_DEBUG, "Returning %d/%d delayed samples.\n",
+                                frame->nb_samples, s->flush_samples);
+        s->flush_samples -= frame->nb_samples;
+    } else {
+        // Trim off samples from the start to compensate for extra decoder
+        // delay. We could also just adjust the pts, but this avoids
+        // including the extra samples in the output altogether.
+        if (s->delay_samples) {
+            int drop_samples = FFMIN(s->delay_samples, frame->nb_samples);
+            av_log(s, AV_LOG_DEBUG, "Dropping %d/%d delayed samples.\n",
+                                    drop_samples, s->delay_samples);
+            s->delay_samples  -= drop_samples;
+            frame->nb_samples -= drop_samples;
+            input_offset = drop_samples * avctx->channels;
+            if (frame->nb_samples <= 0)
+                return 0;
+        }
+    }
+#endif
+
     if ((ret = ff_get_buffer(avctx, frame, 0)) < 0)
         goto end;
 
-    if (frame->pts != AV_NOPTS_VALUE)
-        frame->pts -= av_rescale_q(s->output_delay,
-                                   (AVRational){1, avctx->sample_rate},
-                                   avctx->time_base);
-
-    memcpy(frame->extended_data[0], s->decoder_buffer,
-           avctx->channels * avctx->frame_size *
+    memcpy(frame->extended_data[0], s->decoder_buffer + input_offset,
+           avctx->channels * frame->nb_samples *
            av_get_bytes_per_sample(avctx->sample_fmt));
 
     *got_frame_ptr = 1;
@@ -432,7 +478,11 @@  const AVCodec ff_libfdk_aac_decoder = {
     .decode         = fdk_aac_decode_frame,
     .close          = fdk_aac_decode_close,
     .flush          = fdk_aac_decode_flush,
-    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF,
+    .capabilities   = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_CHANNEL_CONF
+#if FDKDEC_VER_AT_LEAST(2, 5) // 2.5.10
+                      | AV_CODEC_CAP_DELAY
+#endif
+    ,
     .priv_class     = &fdk_aac_dec_class,
     .caps_internal  = FF_CODEC_CAP_INIT_THREADSAFE |
                       FF_CODEC_CAP_INIT_CLEANUP,