On Wed, 05 Jun 2019 22:48:08 +0200 Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx> wrote: > Hi, > > Le mardi 04 juin 2019 à 10:16 +0200, Thierry Reding a écrit : > > On Mon, Jun 03, 2019 at 05:37:11PM +0200, Boris Brezillon wrote: > > > On Mon, 3 Jun 2019 16:05:26 +0200 > > > Thierry Reding <thierry.reding@xxxxxxxxx> wrote: > > > > > > > On Mon, Jun 03, 2019 at 02:51:13PM +0200, Boris Brezillon wrote: > > > > > +Maxime > > > > > > > > > > Oops, just realized Maxime was not Cc-ed on this series. > > > > > > > > > > On Mon, 3 Jun 2019 14:30:20 +0200 > > > > > Thierry Reding <thierry.reding@xxxxxxxxx> wrote: > > > > > > > > > > > On Mon, Jun 03, 2019 at 01:09:42PM +0200, Boris Brezillon wrote: > > > > > > > Some stateless decoders don't support per-slice decoding (or at least > > > > > > > not in a way that would make them efficient or easy to use). > > > > > > > Let's expose a menu to control and expose the supported decoding modes. > > > > > > > Drivers are allowed to support only one decoding but they can support > > > > > > > both too. > > > > > > > > > > > > > > Signed-off-by: Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> > > > > > > > --- > > > > > > > .../media/uapi/v4l/ext-ctrls-codec.rst | 42 ++++++++++++++++++- > > > > > > > drivers/media/v4l2-core/v4l2-ctrls.c | 9 ++++ > > > > > > > include/media/h264-ctrls.h | 13 ++++++ > > > > > > > 3 files changed, 63 insertions(+), 1 deletion(-) > > > > > > > > > > > > > > diff --git a/Documentation/media/uapi/v4l/ext-ctrls-codec.rst b/Documentation/media/uapi/v4l/ext-ctrls-codec.rst > > > > > > > index 82547d5de250..188f625acb7c 100644 > > > > > > > --- a/Documentation/media/uapi/v4l/ext-ctrls-codec.rst > > > > > > > +++ b/Documentation/media/uapi/v4l/ext-ctrls-codec.rst > > > > > > > @@ -1748,6 +1748,14 @@ enum v4l2_mpeg_video_h264_hierarchical_coding_type - > > > > > > > * - __u32 > > > > > > > - ``size`` > > > > > > > - > > > > > > > + * - __u32 > > > > > > > + - ``start_byte_offset`` > > > > > > > + - Where the slice payload starts in the output buffer. Useful when > > > > > > > + operating in per frame decoding mode and decoding multi-slice content. > > > > > > > + In this case, the output buffer will contain more than one slice and > > > > > > > + some codecs need to know where each slice starts. Note that this > > > > > > > + offsets points to the beginning of the slice which is supposed to > > > > > > > + contain an ANNEX B start code > > > > > > > * - __u32 > > > > > > > - ``header_bit_size`` > > > > > > > - > > > > > > > @@ -1931,7 +1939,10 @@ enum v4l2_mpeg_video_h264_hierarchical_coding_type - > > > > > > > - > > > > > > > * - __u16 > > > > > > > - ``num_slices`` > > > > > > > - - Number of slices needed to decode the current frame > > > > > > > + - Number of slices needed to decode the current frame/field. When > > > > > > > + operating in per-slice decoding mode (see > > > > > > > + :c:type:`v4l2_mpeg_video_h264_decoding_mode`), this field > > > > > > > + should always be set to one > > > > > > > * - __u16 > > > > > > > - ``nal_ref_idc`` > > > > > > > - NAL reference ID value coming from the NAL Unit header > > > > > > > @@ -2022,6 +2033,35 @@ enum v4l2_mpeg_video_h264_hierarchical_coding_type - > > > > > > > - 0x00000004 > > > > > > > - The DPB entry is a long term reference frame > > > > > > > > > > > > > > +``V4L2_CID_MPEG_VIDEO_H264_DECODING_MODE (enum)`` > > > > > > > + Specifies the decoding mode to use. Currently exposes per slice and per > > > > > > > + frame decoding but new modes might be added later on. > > > > > > > + > > > > > > > + .. note:: > > > > > > > + > > > > > > > + This menu control is not yet part of the public kernel API and > > > > > > > + it is expected to change. > > > > > > > + > > > > > > > +.. c:type:: v4l2_mpeg_video_h264_decoding_mode > > > > > > > + > > > > > > > +.. cssclass:: longtable > > > > > > > + > > > > > > > +.. flat-table:: > > > > > > > + :header-rows: 0 > > > > > > > + :stub-columns: 0 > > > > > > > + :widths: 1 1 2 > > > > > > > + > > > > > > > + * - ``V4L2_MPEG_VIDEO_H264_DECODING_PER_SLICE`` > > > > > > > + - 0 > > > > > > > + - The decoding is done per slice. v4l2_ctrl_h264_decode_params->num_slices > > > > > > > + must be set to 1 and the output buffer should contain only one slice. > > > > > > > > > > > > I wonder if we need to be that strict. Wouldn't it be possible for > > > > > > drivers to just iterate over a number of slices and decode each in turn > > > > > > if userspace passed more than one? > > > > > > > > > > > > Or perhaps a decoder can batch queue a couple of slices. I'm not sure > > > > > > how frequent this is, but consider something like a spike in activity on > > > > > > your system, causing some slices to get delayed so you actually get a > > > > > > few buffered up before you get a chance to hand them to the V4L2 device. > > > > > > Processing them all at once may help conceal the lag. > > > > > > > > > > Hm, so we'd be in some kind of slice-batch mode, which means we could > > > > > trigger a decode operation with more than one slice, but not > > > > > necessarily all the slices needed to decode a frame. TBH, supporting > > > > > per-frame (or the batch mode you suggest) on a HW that supports > > > > > per-slice decoding should be pretty simple and has not real impact on > > > > > perfs (as you said, it's just a matter of iterating over all the slices > > > > > attached to a decode operation), so I'm fine relaxing the rule here and > > > > > patching the cedrus driver accordingly (I can't really test the > > > > > changes though). Paul, Maxime, what's your opinion? > > So perhaps we could just allow passing any number of slices with each > request within a frame boundary and have userspace set the > V4L2_BUF_FLAG_M2M_HOLD_CAPTURE_BUF flag accordingly. > > I'm not totally sure we need a batching method beyond that, just > submitting requests with groups of slices and the same destination > buffer should work out fine. We can leave it up to the drivers to > handle multiple slices submitted at once. The batching mode we're talking about is exactly that: we let drivers collect slices and send them through a single decode operation. > > > > > We could perhaps have a test program to orchestrate such a scenario. I > > > > think the assumption should obviously still be that we don't cross the > > > > frame boundary using slices in one batch. > > > > > > We should definitely forbid mixing slices of different frames in the > > > same decode operation, since each decode operation is targeting a > > > single capture buffer. > > Agreed. > > > > > Just that if a frame was made > > > > up of, say, 4 slices and you first pass 3 slices, then 1, that it'd be > > > > nice if the driver would be able to cope with that. > > > > > > Yep, that makes sense. > > > > > > > It's something that > > > > could probably even be implemented in the framework as a helper, though > > > > I suspect it'd be just a couple of lines of extra code to wrap a loop > > > > around everything. > > > > > > I also thought about providing generic wrappers, both for this case and > > > the per-slice -> per-frame case (this one would be a bit more > > > complicated as it implies queuing slices in a bounce buffer and > > > triggering the decode operation only when we have all slices of a > > > frame). > > > > I like deferring the addition of that kind of helper until a clear > > pattern emerges out of the drivers that need this, just because that > > gives us real examples on which to model those helpers. But yeah, I > > think it should be possible to have helpers for these for most cases. > > So it seems that we need some convenient way to iterate over each slice > and configure registers accordingly. Perhaps we could have some common > per-codec helpers with callbacks to set the common controls and iterate > over the per-slice controls. Yep, I'm working on extracting common bits out of cedrus/hantro drivers, that new helper (to iterate over all slices) can be part of this series. > > That would help keep our drivers tidy and understandable, and maybe we > could have start/stop steps too, pretty much like we do in cedrus. > > What do you think? Makes total sense, and I'm all for having generic code provided as vb2-m2m/codec helpers.
