Re: [PATCH v3 1/2] virtio-video: Add virtio video device specification

[Keiichi Watanabe <keiichiw@chromium.org>]

Hi,
Thanks Alex for sharing your idea.

On Mon, Apr 6, 2020 at 6:32 PM Alexandre Courbot <acourbot@chromium.org> wrote:
>
> On Mon, Mar 30, 2020 at 6:54 PM Dmitry Sepp <dmitry.sepp@opensynergy.com> wrote:
> >
> > Hi Keiichi,
> >
> > On Freitag, 27. MÃrz 2020 04:35:13 CEST Keiichi Watanabe wrote:
> > > Hi Dmitry,
> > >
> > > On Wed, Mar 25, 2020 at 6:47 PM Dmitry Sepp <dmitry.sepp@opensynergy.com>
> > wrote:
> > > > Hi Keiichi,
> > > >
> > > > On Montag, 23. MÃrz 2020 16:48:14 CET Keiichi Watanabe wrote:
> > > > > Hi,
> > > > >
> > > > > On Mon, Mar 23, 2020 at 10:28 PM Dmitry Sepp
> > > > >
> > > > > <dmitry.sepp@opensynergy.com> wrote:
> > > > > > Hi Keiichi,
> > > > > >
> > > > > > On Montag, 23. MÃrz 2020 13:07:54 CET Keiichi Watanabe wrote:
> > > > > > > Hi everyone,
> > > > > > >
> > > > > > > I have implemented a virtio-video device following my v3 spec in
> > > > > > > crosvm, which worked well together with Dmitry's driver [1]. I've
> > > > > > > started preparing v4 proposal to address problems found while
> > > > > > > implementing the driver and the devices.
> > > > > >
> > > > > > Great news!
> > > > > >
> > > > > > > Regarding v3 protocol, I'm thinking about how we can differentiate
> > > > > > > 'parameters' and 'controls' in the virtio-video spec?
> > > > > > > In the previous discussion, we decided to have a profile, level and
> > > > > > > bitrate as controls because we want to query supported values for
> > > > > > > each
> > > > > > > field.
> > > > > > > But, I don't think it's a good criteria because it'd be possible to
> > > > > > > query other values in params.
> > > > > >
> > > > > > Could you elaborate on this? Do you now how the design could look like
> > > > > > or
> > > > > > it is just an idea? AFAIR during the discussion of OpenSynergy's
> > > > > > original
> > > > > > v1 spec your point was to separate something that we call 'controls'
> > > > > > now
> > > > > > to reduce the command data size and make command handling less error
> > > > > > prone.
> > > > >
> > > > > The problem in v3 is that if we want to add a new value to be set it'd
> > > > > be unclear which params or controls is better to be extended.
> > > > > One possible rule may be "if a value can be queried by the driver, it
> > > > > should be a control". However, this rule doesn't explain why we have
> > > > > "format" in params for example. So, I think we need a discussion and
> > > > > may want to rearrange the structures.
> > > > >
> > > > > Yeah, in the previous discussion, I suggested to have profile, level
> > > > > and bitrate as control values instead of members of params. Now, I'm
> > > > > wondering whether we can have every values as control values.
> > > > > I don't think it's a perfect idea, but I haven't come up with any
> > > > > better concrete design yet. So, I'd really appreciate if you could
> > > > > share your thoughts.
> > > >
> > > > Ok, we can for example add more precise definition to input and output.
> > > > Let's say we have 'bitstream' format structure and a 'image' format
> > > > structure. E.g. for decoder obviously bitstream is input and image is
> > > > output.
> > > >
> > > > Then instead of params and controls we can define some abstract
> > > > 'properties'. And make some of the properties assigned/mapped/available
> > > > to bitstream and some to image, depending on the current function. I
> > > > think that could make sense as for example for decoder 'bitstream'
> > > > probably requires very few basic 'properties' like fourcc format, in
> > > > contrast to 'image'. But for encoder 'bitstream' will also have the
> > > > bitrate 'property' set.
> > >
> > > Ah, it sounds like a good idea to have separate structs for bitstreams
> > > and images.
> > > Okay, let me sort out properties based on the idea.
> > >
> > > # Bitstream
> > > * format
> > > * min/max number of buffers
> > > * bitrate (encoder only)
> > Also depending on format?
> >
> > > * profile (depending on format)
> > > * level (depending on format/profile)
> > >
> > > # Image
> > > * format
> > > * min/max number of buffers
> > > * width/height
> > > * crop information
> > > * number of planes
> > > * plane format
> > > * plane layout
> > * frame rate?
> >
> > >
> > > Then, we have three categories here:
> > > (a) Mandatory properties for bitstreams for both functions
> > > (b) Mandatory properties for images for both functions
> > > (c) Optional properties for bitstream (e.g. bitrate, profile, level)
> > >
> > > So, how about defining structs for each (a), (b) and (c)?
> > >
> > > (a) and (b) can be similar to virtio_video_params in v3 spec draft:
> > > e.g.
> > > struct virtio_video_{bitstream, image}_info {
> > >   int format;
> > >   int min_buffers;
> > >   int max_buffers;
> > >   ...
> > > }
> > Yep, this could make sense. Not sure to which of the two the frame rate should
> > belong. I'd guess to the image.
> >
> > We should also thing of a way to set this. Would we have a set/
> > get_bistream_info() and set/get_image_info()? Plus one generic query/get/
> > set_properties (can use a flexible array to provide several in one step).
> >
> > Or probably event better if properties would apply to the whole stream, not to
> > bitstream or image in particular. WDYT?
>
> Sorry for chiming in late.
>
> Merging the controls and parameters into a single class makes sense to
> me. Both controls and parameters are values that can be updated by
> either the driver or the device, have a default value, and for some of
> them are only relevant depending on the value of other controls (e.g.
> profile depends on the active codec).
>
> V4L2 needs to differentiate between the two because on top of codecs
> it has to be able to manage other kinds of media devices. Since we are
> focused on codecs, we don't have this constraint and thus can probably
> keep things a bit more unified.
>
> How about merging all the properties/controls into a single structure
> which semantics would morph depending on the values of some of its
> members? E.g.

It seems that you're suggesting a big struct where bitstream params,
image params and all controls are merged into, right?
After rethinking, it makes sense to me.
Originally, we had two motivations to have input params and output
params separately:
(1) We could reuse virtio_video_params for both queues.
(2) We have considered extending virtio-video for video capture like V4L2.
However, (1) doesn't seem to be a good idea as per the recent discussion.
Regarding (2), the virtio-video design shouldn't be affected by other
possible use cases. They should be discussed after we finalize
virtio-video design.
Hence, I guess we have no reason to stick to separate params into two.
I'm supportive of this idea now.
Also, having only one struct would simplify updating parameters.

>
> struct codec_params {
>     u32 frame_width;
>     u32 frame_height;
>     /* H264, VP8, VP9, etc */
>     enum virtio_codec format;
>     union {
>         struct {
>             enum h264_profile profile;
>             enum h264_level level;
>         } h264;
>         struct {
>             enum vp8_profile profile;
>         } vp8;
>         ....
>     };
>     u32 bitrate;
>     ....
> };

More specifically, this struct would be:

// the word "codec" might not be needed in the struct name.
struct virtio_video_codec_params {
    // Image format
    enum virtio_video_frame_format frame_format;
    le32 frame_width;
    le32 frame_height;
    le32 min_frame_buffers;
    le32 max_frame_buffers;
    le32 cur_frame_buffers; // It's needed for REQBUFS's "count"
    struct virtio_video_crop crop;
    le32 frame_rate;
    le32 num_planes;
    struct virtio_video_plane_format \
        plane_formats[VIRTIO_VIDEO_MAX_PLANES];

    // Bitstream format
    enum virtio_video_coded_format coded_format;
    le32 min_coded_buffers;
    le32 max_coded_buffers;
    le32 cur_coded_buffers;
    le32 bitrate;
    union {
        struct {
          enum h264_profile profile;
          enum h264_level level;
        } h264;
        struct {
          enum vp8_profile profile;
        } vp8;
        ...
    } codec;
}


>
> The idea being that depending on the value of "format", only the
> relevant member of the union becomes meaningful. This ensures that the
> device/driver does not need to check for every control whether it is
> valid in the current context ; it just needs to check what "format" is
> and take the values from the relevant members.

I like this idea to use union to make it more structured.

>
> I don't think we even need to have a different queue for both structs
> (which is what V4L2 does, but again for its own reasons). Just a
> single one per coding or decoding context could be enough AFAICT.
>
> Depending on whether we are decoding or encoding, access to some
> members would be restricted to read/only for the device or driver. For
> instance, when encoding the driver can set "bitrate" to the desired
> encoding rate. When decoding, the decoder can report the video's
> bitrate there.
>
> Now I'm not sure what would be the best way to share this structure.
> Either a memory area shared between the driver and device, with
> commands/messages sent to notify that something has changed, or
> sending the whole structure with each command/message.

I don't think the first idea is feasible in the virtio mechanism. So,
we can utilize the same way as the previous proposal; SET_PARAMS and
GET_PARAMS.

We also need to think of how to advertise supported profiles and
levels. Probably, we might want to extend struct
virtio_video_format_desc to include supported profiles/levels in a
response of QUERY_CAPABLITY.

>
> Now the parameters I have listed above are not subject to changing a
> lot, but there are also parameters that we may want to specify/be
> notified on with each frame. For instance, whether we want a frame to
> be forcibly encoded as a keyframe. V4L2 uses a control for this, but
> we could probably do better if we can pass this information with each
> frame to be encoded. Maybe we can implement that by using different
> QUEUE commands for encoder and decoder, or again by using a union.

Ah, I haven't come up with such a kind of parameter. Perhaps, we can
extend struct virtio_video_resource_queue to have this flag.

Best regards,
Keiichi

>
> >
> > Best regards,
> > Dmitry.
> >
> > >
> > > (c) would be very similar to struct virtio_video_*_control in v3.
> > > Renaming them to 'properties' would be a nice idea as Dmitry said.
> > >
> > > While the designs of structs are not changed from 'params' and
> > > 'controls', we now have rules for differentiation at least.
> > > What do you think?
> >
> > >
> > > Best regards,
> > > Keiichi
> > >
> > > > > > On one hand if don't really see any difference in params vs controls
> > > > > > it
> > > > > > would for sure make sense to remove one of the two. On the other hand
> > > > > > I'd
> > > > > > of course like to avoid moving back in forth, especially when it comes
> > > > > > to
> > > > > > such a major driver rework.
> > > > >
> > > > > Yes, I understand that it may require a big change in the
> > > > > implementation.
> > > > > I'm sorry for not being able to think of this point seriously in the
> > > > > previous thread.
> > > > >
> > > > > Of course, I'd also really like to avoid rework, but I believe we
> > > > > shouldn't give up defining a clean and reasonable specification.
> > > > > Let's find a clear definition in this cycle to avoid future rework.
> > > > >
> > > > > > > So, I'm thinking about what should be the difference between params
> > > > > > > and controls. If no difference, we should deprecate
> > > > > > > virtio_video_params and have every field there as a control value
> > > > > > > instead.
> > > > > >
> > > > > > I deem we should then deprecate controls instead. Params seem to be
> > > > > > more
> > > > > > abstract. Width and height don't sound like a control for me.
> > > > >
> > > > > Though this is actually one of options, how can we query profiles and
> > > > > levels if they are in params?
> > > > > This is why we decided them as controls.
> > > > >
> > > > > Best regards,
> > > > > Keiichi
> > > > >
> > > > > > > If we add a new command to get and set multiple controls at once,
> > > > > > > this
> > > > > > > change won't cause overhead.
> > > > > >
> > > > > > How would we do this? Provide a flexible array member where each entry
> > > > > > has
> > > > > > a type field first?
> > > > >
> > > > > Yeah, something like the idea. But, I haven't designed an actual
> > > > > structure
> > > > > yet.
> > > > >
> > > > > > What can also make sense is to potentially join set and get calls
> > > > > > (probably
> > > > > > provide 'get' stuff automatically within a response to 'set'). Anyway
> > > > > > set
> > > > > > and get are currently used in conjunction all the time.
> > > > >
> > > > > It'd make sense to return new input and output params when one of
> > > > > params is updated.
> > > > > But, if we choose this design, we need to assume one device has just
> > > > > two params; input and output.
> > > > >
> > > > > This is okay for video decoder and encoder, but it may become a
> > > > > problem if we want to support other types of video device that has
> > > > > only one direction. (e.g. video capture device)
> > > > > Though we have no plan for supporting this, OpenSynergy's v1 proposal
> > > > > contained this type IIRC.
> > > >
> > > > Honestly speaking, the idea is not completely abandoned. The spec and the
> > > > driver has more than enough functionality to handle a simple Android EVS
> > > > camera use-case. But I think let's discuss this separately later.
> > > >
> > > > Best regards,
> > > > Dmitry.
> > > >
> > > > > Best regards,
> > > > > Keiichi
> > > > >
> > > > > > Best regards,
> > > > > > Dmitry.
> > > > > >
> > > > > > > What do you think? Is there anything I missed?
> > > > > > > If it sounds fine, I'll remove virtio_video_params from the v4 spec
> > > > > > > proposal.
> > > > > > >
> > > > > > > Best regards,
> > > > > > > Keiichi
> > > > > > >
> > > > > > > [1]: https://patchwork.linuxtv.org/patch/61717/
> >
> >