[FFmpeg-devel] [PATCH v3 3/3] lavc/libdavs2.c: use decoder data directly instead of memcpy

[hwren]

I'm sorry I almost missed this email. Thanks for your review.

At 2019-07-23 03:42:20, "Marton Balint" <cus at passwd.hu> wrote:
>
>
>On Sat, 13 Jul 2019, hwren wrote:
>
>>
>> At 2019-07-13 01:38:55, "Marton Balint" <cus at passwd.hu> wrote:
>>>
>>>
>>> On Fri, 12 Jul 2019, hwrenx wrote:
>>>
>>>> Can effectivly improved decoding speed when memcpy becomes a limitation
>>>> for proccessing high resolution source.
>>>> Tested under i7-8700k with `ffmpeg -i 7680x4320.avs2 -vsync 0 -f null -`
>>>> got performance 23fps => 42fps
>>>>
>>>> Signed-off-by: hwrenx <hwrenx at 126.com>
>>>> ---
>>>> libavcodec/libdavs2.c | 54 +++++++++++++++++++++++++++++----------------------
>>>> 1 file changed, 31 insertions(+), 23 deletions(-)
>>>>
>>>> diff --git a/libavcodec/libdavs2.c b/libavcodec/libdavs2.c
>>>> index 1b274a3..af0778f 100644
>>>> --- a/libavcodec/libdavs2.c
>>>> +++ b/libavcodec/libdavs2.c
>>>> @@ -60,13 +60,24 @@ static av_cold int davs2_init(AVCodecContext *avctx)
>>>>     return 0;
>>>> }
>>>> 
>>>> +static void davs2_frame_unref(void *opaque, uint8_t *data) {
>>>> +    DAVS2Context    *cad = (DAVS2Context *)opaque;
>>>> +    davs2_picture_t  pic;
>>>> +
>>>> +    pic.magic = (davs2_picture_t *)data;
>>>> +
>>>> +    if (cad->decoder) {
>>>> +        davs2_decoder_frame_unref(cad->decoder, &pic);
>>>> +    } else {
>>>> +        av_log(NULL, AV_LOG_WARNING, "Decoder not found, frame unreference failed.\n");
>>>
>>> Whoa, this should not happen, and you have to be prepared that the user 
>>> might close the decoder before freeing the last frame.
>>
>> I hope so, actually it's harmless if we failed to unref some frames outside davs2 when decoder
>> was closed through davs2_decoder_close(). The decoder will free the memory which contains
>> all recycle data itself.
>
>And that is the problem, because the data should survive closing the 
>(avcodec) decoder.

Do you mean after decoder closing, there may still some data left which should be unref by decoder ...
The davs2_decoder_frame_unref() can only notice the decoder that it could reuse this part of memory,
but the decoder won't free or realloc any memory. All these memory will be free together when decoder
was closing. So the "survived data" in ffmpeg should just be some unuseful pointers.

Or you mean even decoder is closed, there are still some data needed by ffmpeg?

>
>>
>>>
>>> Maybe you should use some refcounting and create references to the 
>>> decoder, others may have a better idea.
>>>
>>> Regards,
>>> Marton
>>
>> Of course, anyway, it would be better if there are any ways to make sure we correctly freed all
>> frames even decoder was not found or exited without a davs2_decoder_close().  But force to
>> free the memory immediately in the call back function can cause troubles to decoder and I'm
>> not very clearly understand how to use reference to deal with this problem inside a codec.
>>
>>
>> Maybe that means I could reference the buffer and waiting ffmpeg to free them once they were
>> not handled by davs2?
>
>I am not sure I understand your concern here.
>
>Here is what I propose in more detail, I think it works for all cases:
>
>In davs_init allocate a small struct which contains the decoder and a 
>reference counter for the decoder itself:
>DAVS2Reference {
>   atomic_int refcount;
>   void *decoder;
>}
>set the refs to 1.
>
>In davs2_dump_frames you should pass the allocated DAVS2Reference to 
>av_buffer_create instead of the DAVS2Context. After each successful 
>av_buffer_create you should increase refcount.
>
>In davs2_frame_unref after freeing the frame you should decrease the 
>reference counter and if you reach 0 then you can destroy the davs2 
>decoder (and free DAVS2Reference struct as well)
>
>You can do the same in davs2_end: decrease the refcount and if you reach 0 
>then you can close the davs2 decoder and free the DAVS2Reference struct.
>

In my understanding, once closing decoder, ffmpeg should finish or stop outputing (not so sure,
if not please correct me). So I think normally the refcount should always be zero when using
decoder_close(), if not zero, maybe decoder can do nothing but wait (and never stop).

If after closing decoder there are still some data needed by ffmpeg, a refcounter is necessary indeed.

Regards,
Huiwen REN

>Regards,
>Marton
>_______________________________________________
>ffmpeg-devel mailing list
>ffmpeg-devel at ffmpeg.org
>https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>To unsubscribe, visit link above, or email
>ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".