avcodec_flush_buffers() not resetting E-AC-3 decoder

[Peter Krefting]

Summary of the bug:
I am reusing the same AVCodecContext to decode several parallel E-AC-3 audio streams. As per documentation I am calling avcodec_flush_buffers() between each decoding run. This seems to work for most codecs, but when used with the E-AC-3 decoder the output is not identical between runs

How to reproduce:

#include <libavcodec/avcodec.h>
#include <libavformat/avformat.h>
#include <stdio.h>
#include <assert.h>

int main(int argc, char *argv[])
{
    const AVCodec *codec = avcodec_find_decoder(AV_CODEC_ID_EAC3);
    AVCodecContext *decoder = avcodec_alloc_context3(codec);
    avcodec_open2(decoder, codec, NULL);
    assert(decoder->sample_fmt == AV_SAMPLE_FMT_FLTP);

    for (int l = 0; l < 2; ++ l)
    {
        printf("[%d] Reading file\n", l);
        avcodec_flush_buffers(decoder);
        AVFormatContext *avf_ctx = NULL;
        avformat_open_input(&avf_ctx, "file:i28553p303t610156395819694.ac3", NULL, NULL);
        avformat_find_stream_info(avf_ctx, NULL);

        AVPacket avpkt = { 0 };
        while (av_read_frame(avf_ctx, &avpkt) == 0)
        {
            avcodec_send_packet(decoder, &avpkt);
            AVFrame *frame = av_frame_alloc();
            avcodec_receive_frame(decoder, frame);
            printf("Got %d samples in %d channels\n", frame->nb_samples, frame->channels);

            for (int i = 0; i < frame->nb_samples; ++ i)
            {
                printf("%4d:", i);
                for (int j = 0; j < frame->channels; ++ j)
                {
                    printf(" [%d]%f", j, ((float *) frame->data[j])[i]);
                }
                printf("\n");
            }

            av_frame_free(&frame);
        }
        avformat_close_input(&avf_ctx);
    }

    avcodec_free_context(&decoder);
    return 0;
}

The test file is available here: ​https://e.pcloud.link/publink/show?code=XZS8y1ZlEryJIDYU0yIzDTUhw1sDfHDRtUV and contains a ten-second extract of an audio track from an MPEGTS container, captured off-air while looking at a problem where the encoder suddenly only outputs silence.

I expect the output to be identical on the first and second run, but that is not what I am seeing:

$ ./decode |grep -A5 'Reading file'
[eac3 @ 0x55fc4b0a5ec0] Estimating duration from bitrate, this may be inaccurate
[0] Reading file
Got 1536 samples in 6 channels
   0: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
   1: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
   2: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
   3: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
[eac3 @ 0x55fc4b11a840] Estimating duration from bitrate, this may be inaccurate
--
[1] Reading file
Got 1536 samples in 6 channels
   0: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
   1: [0]-0.000000 [1]0.000000 [2]-0.000000 [3]-0.000000 [4]-0.000000 [5]-0.000000
   2: [0]0.000000 [1]0.000000 [2]0.000000 [3]0.000000 [4]0.000000 [5]0.000000
   3: [0]-0.000000 [1]-0.000000 [2]-0.000000 [3]-0.000000 [4]-0.000000 [5]-0.000000

In real life we are also interleaving other E-AC-3 audio streams from the same MPEGTS between decodes, and see that the initial decoded samples differ even more than when just redecoding the same sample over and over.

[Peter Krefting]

We are also seeing a similar issue in an unrelated case using the HEVC decoder (AV_CODEC_ID_H265). When multiplexing several streams into the same AVCodecContext object, calling avcodec_flush_buffers() between each call, we sometimes fail to get frames decoded, but if we change to having a dedicated AVCodecContext for each of the streams it decodes fine.

[Peter Krefting]

Also verified that the I see the same issue in the 6.1 release version. Updating the build framework is fairly heavy process, and I have not tested the latest git master.