Re: [PATCH v3 0/2] media: cedrus: Add H264 decoding support

[Maxime Ripard]

On Wed, Feb 13, 2019 at 01:28:34PM -0300, Ezequiel Garcia wrote:
> On Wed, 2019-02-13 at 12:02 +0900, Tomasz Figa wrote:
> > On Wed, Feb 13, 2019 at 6:22 AM Ezequiel Garcia <ezequiel@xxxxxxxxxxxxx> wrote:
> > > Hey Tomasz,
> > > 
> > > On Tue, 2019-02-12 at 21:50 +0900, Tomasz Figa wrote:
> > > > Hi Maxime,
> > > > 
> > > > On Mon, Feb 11, 2019 at 11:39 PM Maxime Ripard
> > > > <maxime.ripard@xxxxxxxxxxx> wrote:
> > > > > Hi,
> > > > > 
> > > > > Here is a new version of the H264 decoding support in the cedrus
> > > > > driver.
> > > > 
> > > > Thanks for working on this. Please see my comments below.
> > > > 
> > > > > As you might already know, the cedrus driver relies on the Request
> > > > > API, and is a reverse engineered driver for the video decoding engine
> > > > > found on the Allwinner SoCs.
> > > > > 
> > > > > This work has been possible thanks to the work done by the people
> > > > > behind libvdpau-sunxi found here:
> > > > > https://github.com/linux-sunxi/libvdpau-sunxi/
> > > > > 
> > > > > I've tested the various ABI using this gdb script:
> > > > > http://code.bulix.org/jl4se4-505620?raw
> > > > > 
> > > > > And this test script:
> > > > > http://code.bulix.org/8zle4s-505623?raw
> > > > > 
> > > > > The application compiled is quite trivial:
> > > > > http://code.bulix.org/e34zp8-505624?raw
> > > > > 
> > > > > The output is:
> > > > > arm:    builds/arm-test-v4l2-h264-structures
> > > > >         SHA1: 88cbf7485ba81831fc3b93772b215599b3b38318
> > > > > x86:    builds/x86-test-v4l2-h264-structures
> > > > >         SHA1: 88cbf7485ba81831fc3b93772b215599b3b38318
> > > > > x64:    builds/x64-test-v4l2-h264-structures
> > > > >         SHA1: 88cbf7485ba81831fc3b93772b215599b3b38318
> > > > > arm64:  builds/arm64-test-v4l2-h264-structures
> > > > >         SHA1: 88cbf7485ba81831fc3b93772b215599b3b38318
> > > > > 
> > > > > Let me know if there's any flaw using that test setup, or if you have
> > > > > any comments on the patches.
> > > > > 
> > > > > Maxime
> > > > > 
> > > > > Changes from v2:
> > > > >   - Simplified _cedrus_write_ref_list as suggested by Jernej
> > > > >   - Set whether the frame is used as reference using nal_ref_idc
> > > > >   - Respect chroma_format_idc
> > > > >   - Fixes for the scaling list and prediction tables
> > > > >   - Wrote the documentation for the flags
> > > > >   - Added a bunch of defines to the driver bit fields
> > > > >   - Reworded the controls and data format descriptions as suggested
> > > > >     by Hans
> > > > >   - Reworked the controls' structure field size to avoid padding
> > > > >   - Removed the long term reference flag
> > > > 
> > > > This and...
> > > > 
> > > 
> > > Maxime has dropped this because of Ayaka's mail about long term references
> > > not making much sense in stateless decoders.
> > 
> > I haven't seen any argument confirming that thesis, though. I should
> > have kicked in earlier, sorry.
> > 
> 
> OK, in that case, we need to have this flag back.
> 
> > > I noticed that RK3399 TRM has a field to specify long term refs and
> > > so was wondering about this item as well.
> > > 
> > > > >   - Reintroduced the neighbor info buffer
> > > > >   - Removed the ref_pic_list_p0/b0/b1 arrays that are redundant with the
> > > > >     one in the DPB
> > > > 
> > > > these are used in our Rockchip VDEC driver.
> > > > 
> > > > Could you elaborate on the reasons why they got removed?
> > > > 
> > > 
> > > If I understood correctly, there are two reference picture lists.
> > > P-frames will populate ref_pic_list0 and B-frames will populate both.
> > > 
> > > According to this, v4l2_ctrl_h264_slice_param.ref_pic_list0 and .ref_pic_list1
> > > should be enough and ref_pic_list_p0/b0/b1 are not needed.
> > > 
> > > What do you think?
> > 
> > The lists in v4l2_ctrl_h264_slice_param are expected to be past the
> > per-slice modification stage (which is quite complicated and better
> > done in userspace),
> 
> The fact that these are RefPicList0 and RefPicList1, after
> the reordering stage should be better documented.
> 
> > while the ones in v4l2_ctrl_h264_decode_param just
> > in the original order. Rockchip VPU expects them in the original order
> > and does the modification in the hardware.
> > 
> 
> OK, I see.
> 
> So, we have RefPicList0 and RefPicList1, and there is an initialization
> stage and a modification/reordering process.
> 
> One could argue that it's more generic to just pass the initial list,
> but that would mean doing in the kernel something that is easier
> done in userspace (and parsers doing this are already available).
> 
> The question would be what is the most generic way of passing
> the RefPicList0 and RefPicList1 in its initial state.
> 
> 1/ We create additional controls for these.
> 
> 2/ We put them on some of the other controls. Putting them on
> v4l2_ctrl_h264_decode_param didn't seem too wrong.
> 
> Any objections to put them back in there?

None. i'll put them back in.

Maxime

-- 
Maxime Ripard, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com