[FFmpeg-devel] [PATCH] avcodec/mediacodecdec: call MediaCodec.stop on close

[sfan5]

Am 08.08.24 um 17:29 schrieb Zhao Zhili:
>> On Aug 8, 2024, at 00:27, sfan5<sfan5 at live.de>  wrote:
>>
>> Hi all,
>>
>> attached is a small fix for the MediaCodec code. Tested on Android 14.
>>
>> This can free up vital resources in case of using multiple
>> decoding instances and there are buffer references left over
>> and not immediately cleaned up.
>>
>> Signed-off-by: sfan5<sfan5 at live.de>
>> ---
>>   libavcodec/mediacodecdec_common.c | 12 ++++++++++++
>>   1 file changed, 12 insertions(+)
>>
>> diff --git a/libavcodec/mediacodecdec_common.c b/libavcodec/mediacodecdec_common.c
>> index d6f91e6e89..c888dea8cf 100644
>> --- a/libavcodec/mediacodecdec_common.c
>> +++ b/libavcodec/mediacodecdec_common.c
>> @@ -841,6 +841,18 @@ int ff_mediacodec_dec_flush(AVCodecContext *avctx, MediaCodecDecContext *s)
>>   
>>   int ff_mediacodec_dec_close(AVCodecContext *avctx, MediaCodecDecContext *s)
>>   {
>> +    if (!s)
>> +        return 0;
>> +
>> +    if (s->codec) {
>> +        if (atomic_load(&s->hw_buffer_count) == 0) {
>> +            ff_AMediaCodec_stop(s->codec);
>> +            av_log(avctx, AV_LOG_DEBUG, "MediaCodec %p stopped\n", s->codec);
>> +        } else {
>> +            av_log(avctx, AV_LOG_DEBUG, "Not stopping MediaCodec (there are buffers pending)\n");
>> +        }
>> +    }
>> +
> Could you elaborate on how this resolved the issue?
> If hw_buffer_count is zero, isn’t MediaCodecDecContext will be released immediately?
> If hw_buffer_count isn’t zero, the patch make no real change, so how does this patch work?

Sure.

here's the original report: 
https://github.com/mpv-android/mpv-android/issues/966

summary:

mpv's hwdec code uses a single surface to facilitate zero-copy video 
playback. Keeping an active MediaCodec instance connected to the surface 
blocks the same surface from being used with a different MediaCodec 
instance.

However mpv also keeps a reference to the last rendered frame alive even 
when switching between files. It also flushes the codec when it's done 
with a file. This leads to a situation where hw_buffer_count == 0 && 
refcount > 1.

If you were to say that this should be fixed in mpv I would agree and I 
have indeed proposed a PR for this. But I noticed we're never calling 
stop() in mediacodecdec and it's a very simple remedy.