Hi Philipp, Hans, > On mem2mem decoders with a hardware bitstream ringbuffer, to drain the > buffer at the end of the stream, remaining frames might need to be > decoded without additional input buffers being provided, and after > calling streamoff on the v4l2 output queue. This also allows a driver > to copy input buffers into their bitstream ringbuffer and immediately > mark them as done to be dequeued. > > The motivation for this patch is hardware assisted h.264 reordering > support in the coda driver. For high profile streams, the coda can hold > back out-of-order frames, causing a few mem2mem device runs in the > beginning, that don't produce any decompressed buffer at the v4l2 > capture side. At the same time, the last few frames can be decoded from > the bitstream with mem2mem device runs that don't need a new input > buffer at the v4l2 output side. A streamoff on the v4l2 output side can > be used to put the decoder into the ringbuffer draining end-of-stream > mode.
If I remember correctly the main goal of introducing the m2m framework was to support simple mem2mem devices where one input buffer = one output buffer. In other cases m2m was not to be used. An example of such mem2mem driver, which does not use m2m framework is MFC. It uses videobuf2 directly and it is wholly up to the driver how will it control the queues, stream on/off and so on. You can then have one OUTPUT buffer generate multiple CAPTURE buffer, multiple OUTPUT buffers generate a single CAPTURE buffer. Consume OUTPUT buffer without generating CAPTURE buffer (e.g. when starting decoding) and produce CAPTURE buffers without consuming OUTPUT buffers (e.g. when finishing decoding). I think that stream off should not be used to signal EOS. For this we have EOS event. You mentioned the EOS buffer flag. This is the idea originally proposed by Andrzej Hajda, after some lengthy discussions with v4l community this idea was changed to use an EOS event. I was all for the EOS buffer flag, but after discussion with Mauro I understood his arguments. We can get back to this discussion, if we are sure that events are not enough. Please also note that we need to keep backward compatibility. Original EOS buffer flag patches by Andrzej and part of the discussion can be found here: 1) https://linuxtv.org/patch/10624/ 2) https://linuxtv.org/patch/11373/ Best wishes, Kamil Debski -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
