Message-ID: <11988268.O9o76ZdvQC@jernej-laptop>
Date:   Wed, 01 Jun 2022 18:35:23 +0200
From:   Jernej Škrabec <jernej.skrabec@...il.com>
To:     mchehab@...nel.org, ezequiel@...guardiasur.com.ar,
        p.zabel@...gutronix.de, gregkh@...uxfoundation.org,
        mripard@...nel.org, paul.kocialkowski@...tlin.com, wens@...e.org,
        samuel@...lland.org, andrzej.p@...labora.com,
        Hans Verkuil <hverkuil@...all.nl>,
        Benjamin Gaignard <benjamin.gaignard@...labora.com>,
        Nicolas Dufresne <nicolas.dufresne@...labora.com>
Cc:     linux-media@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-rockchip@...ts.infradead.org, linux-staging@...ts.linux.dev,
        linux-arm-kernel@...ts.infradead.org, linux-sunxi@...ts.linux.dev,
        kernel@...labora.com
Subject: Re: [PATCH v6 11/17] media: uapi: Add V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS control

Dne sreda, 01. junij 2022 ob 18:20:53 CEST je Nicolas Dufresne napisal(a):
> Le mardi 31 mai 2022 à 20:20 +0200, Jernej Škrabec a écrit :
> > Dne torek, 31. maj 2022 ob 08:58:46 CEST je Benjamin Gaignard napisal(a):
> > > Le 30/05/2022 à 23:24, Jernej Škrabec a écrit :
> > > > Dne ponedeljek, 30. maj 2022 ob 15:49:57 CEST je Hans Verkuil 
napisal(a):
> > > > > On 30/05/2022 11:18, Hans Verkuil wrote:
> > > > > > On 29/05/2022 08:40, Jernej Škrabec wrote:
> > > > > > > Hi!
> > > > > > > 
> > > > > > > This series looks very good and I plan to test it shortly on
> > > > > > > Cedrus,
> > 
> > but
> > 
> > > > I
> > > > 
> > > > > > > have one major concern below.
> > > > > > > 
> > > > > > > Dne petek, 27. maj 2022 ob 16:31:28 CEST je Benjamin Gaignard
> > 
> > napisal(a):
> > > > > > > > The number of 'entry point offset' can be very variable.
> > > > > > > > Instead of using a large static array define a v4l2 dynamic
> > > > > > > > array
> > > > > > > > of U32 (V4L2_CTRL_TYPE_U32).
> > > > > > > > The number of entry point offsets is reported by the elems
> > > > > > > > field
> > > > > > > > and in struct
> > > > > > > > v4l2_ctrl_hevc_slice_params.num_entry_point_offsets
> > > > > > > > field.
> > > > > > > 
> > > > > > > Slice control by itself is variable length array, so you would
> > > > > > > actually
> > > > 
> > > > need
> > > > 
> > > > > > > 2D variable array for entry points which is not supported.
> > > > > > > However,
> > 
> > easy
> > 
> > > > > > > workaround for that is to flatten 2D array to 1D and either have
> > > > > > > another
> > > > 
> > > > slice
> > > > 
> > > > > > > control field which would tell first entry point index for
> > > > > > > convenience or
> > > > 
> > > > let
> > > > 
> > > > > > > driver calculate it by adding up all num_entry_point_offsets of
> > > > > > > previous
> > > > > > > slices.
> > > > > > > 
> > > > > > > Hans, what do you think?
> > > > > > 
> > > > > > If I would support 2D variable array sizes, wouldn't that be more
> > 
> > elegant?
> > 
> > > > > > The current implementation doesn't support that, but as the commit
> > > > > > log
> > 
> > for
> > 
> > > > > > patch 1/17 says:
> > > > > > 
> > > > > > "Currently dynamically sized arrays are limited to one dimensional
> > 
> > arrays,
> > 
> > > > > > but that might change in the future if there is a need for it."
> > > > > > 
> > > > > > Let me know if you agree, and I'll try to implement this. It's
> > > > > > been a
> > > > 
> > > > while
> > > > 
> > > > > > since I last looked at this, so I'm not sure how much work it is,
> > > > > > but it
> > > > 
> > > > is
> > > > 
> > > > > > probably worth a shot.
> > > > > 
> > > > > Digging more into this made me realize that this doesn't actually
> > > > > help
> > 
> > for
> > 
> > > > this
> > > > 
> > > > > particular case.
> > > > > 
> > > > > I would lean towards your second suggestion of adding up all
> > > > 
> > > > num_entry_point_offsets
> > > > 
> > > > > of previous slices.
> > > > 
> > > > Just one question/clarification about dynamic arrays - does driver
> > > > need to
> > > > reserve maximum amount of memory for dynamic array control at
> > 
> > initialization
> > 
> > > > time? If so, this would still be problematic, since there cound be a
> > > > huge
> > > > amount of entry points in theory.
> > > 
> > > When adding the control the driver could set .dims field to specify
> > > the max number of accepted slices.
> > > I have added '#define V4L2_HEVC_SLICE_MAX_COUNT 600' that you could use
> > > for this field. It is the value we have found when using slices with
> > > RKVDEC
> > > driver.
> > 
> > Is this maximum value applicable only to RKVDEC or is universal? Anyway,
> > this means maximum offset points control for Cedrus would be 600 * 1024
> > (max. offset points supported per slice) * 4 ~= 2.4 MB, which is a lot
> > for one control, but I can live with that...
> 
> I believe its defined following "Table A.8 – General tier and level limits".
> With the assumption there will never be a level 7 (which I think is fair).
> If anyone saw other reasons for this limit, let me know.
> 
> This is a worse case scenario, this is quite unlikely in practice, so while
> performance might be a disaster if your craft a stream for that case, I
> don't think it will ever happen in real life.

But do we really need to cover worst case scenario? In theory, one driver can 
set limit to (for example) max 100 slices and if there is a frame with 600 
slices, userspace app would submit 6 decode requests. Basically the same way 
it's done today. While not as performant, it would be good compromise between 
resources and speed.

> 
> > Best regards,
> > Jernej
> > 
> > > Regards,
> > > Benjamin
> > > 
> > > > Best regards,
> > > > Jernej
> > > > 
> > > > > Regards,
> > > > > 
> > > > > 	Hans
> > > > > 	
> > > > > > Regards,
> > > > > > 
> > > > > > 	Hans
> > > > > > 	
> > > > > > > Note, it seems that H265 decoding on Cedrus still works without
> > > > > > > entry
> > > > 
> > > > points,
> > > > 
> > > > > > > so this problem can be solved later. I'm not sure what we lose
> > > > > > > with
> > 
> > that
> > 
> > > > but
> > > > 
> > > > > > > it was suggested that this could influence speed or error
> > > > > > > resilience or
> > > > 
> > > > both.
> > > > 
> > > > > > > However, I think we're close to solve it, so I'd like to do that
> > > > > > > now.
> > > > > > > 
> > > > > > > Best regards,
> > > > > > > Jernej
> > > > > > > 
> > > > > > > > Signed-off-by: Benjamin Gaignard
> > > > > > > > <benjamin.gaignard@...labora.com>
> > > > > > > > ---
> > > > > > > > 
> > > > > > > >   .../userspace-api/media/v4l/ext-ctrls-codec.rst       | 11
> > > > > > > >   +++++++++
> > 
> > ++
> > 
> > > > > > > >   drivers/media/v4l2-core/v4l2-ctrls-defs.c             |  5
> > > > > > > >   +++++
> > > > > > > >   include/media/hevc-ctrls.h                            |  5
> > > > > > > >   ++++-
> > > > > > > >   3 files changed, 20 insertions(+), 1 deletion(-)
> > > > > > > > 
> > > > > > > > diff --git
> > > > > > > > a/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > 
> > b/
> > 
> > > > > > > Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > > 
> > > > > > > > index 0796b1563daa..05228e280f66 100644
> > > > > > > > ---
> > > > > > > > a/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > > > +++
> > > > > > > > b/Documentation/userspace-api/media/v4l/ext-ctrls-codec.rst
> > > > > > > > @@ -3010,6 +3010,9 @@ enum
> > > > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > > > > 
> > > > > > > >       * - __u32
> > > > > > > >       
> > > > > > > >         - ``data_bit_offset``
> > > > > > > >         - Offset (in bits) to the video data in the current
> > > > > > > >         slice data.
> > > > > > > > 
> > > > > > > > +    * - __u32
> > > > > > > > +      - ``num_entry_point_offsets``
> > > > > > > > +      - Specifies the number of entry point offset syntax
> > > > > > > > elements in
> > 
> > the
> > 
> > > > > > > slice header.
> > > > > > > 
> > > > > > > >       * - __u8
> > > > > > > >       
> > > > > > > >         - ``nal_unit_type``
> > > > > > > >         - Specifies the coding type of the slice (B, P or I).
> > > > > > > > 
> > > > > > > > @@ -3150,6 +3153,14 @@ enum
> > > > > > > > v4l2_mpeg_video_hevc_size_of_length_field -
> > > > > > > > 
> > > > > > > >       \normalsize
> > > > > > > > 
> > > > > > > > +``V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS (integer)``
> > > > > > > > +    Specifies entry point offsets in bytes.
> > > > > > > > +    This control is a dynamically sized array. The number of
> > > > > > > > entry
> > > > 
> > > > point
> > > > 
> > > > > > > > +    offsets is reported by the ``elems`` field.
> > > > > > > > +    This bitstream parameter is defined according to
> > > > > > > > :ref:`hevc`.
> > > > > > > > +    They are described in section 7.4.7.1 "General slice
> > > > > > > > segment
> > 
> > header
> > 
> > > > > > > > +    semantics" of the specification.
> > > > > > > > +
> > > > > > > > 
> > > > > > > >   ``V4L2_CID_STATELESS_HEVC_SCALING_MATRIX (struct)``
> > > > > > > >   
> > > > > > > >       Specifies the HEVC scaling matrix parameters used for
> > > > > > > >       the scaling
> > > > > > > 
> > > > > > > process
> > > > > > > 
> > > > > > > >       for transform coefficients.
> > > > > > > > 
> > > > > > > > diff --git a/drivers/media/v4l2-core/v4l2-ctrls-defs.c
> > > > > > > > b/drivers/media/
> > > > 
> > > > v4l2-
> > > > 
> > > > > > > core/v4l2-ctrls-defs.c
> > > > > > > 
> > > > > > > > index d594efbcbb93..e22921e7ea61 100644
> > > > > > > > --- a/drivers/media/v4l2-core/v4l2-ctrls-defs.c
> > > > > > > > +++ b/drivers/media/v4l2-core/v4l2-ctrls-defs.c
> > > > > > > > @@ -1188,6 +1188,7 @@ const char *v4l2_ctrl_get_name(u32 id)
> > > > > > > > 
> > > > > > > >   	case V4L2_CID_STATELESS_HEVC_DECODE_PARAMS:		
return
> > > > > > > 
> > > > > > > "HEVC Decode Parameters";
> > > > > > > 
> > > > > > > >   	case V4L2_CID_STATELESS_HEVC_DECODE_MODE:		
return
> > > > > > > 
> > > > > > > "HEVC Decode Mode";
> > > > > > > 
> > > > > > > >   	case V4L2_CID_STATELESS_HEVC_START_CODE:		
return
> > > > > > > 
> > > > > > > "HEVC Start Code";
> > > > > > > 
> > > > > > > > +	case V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS:	
return
> > > > > > > 
> > > > > > > "HEVC Entry Point Offsets";
> > > > > > > 
> > > > > > > >   	/* Colorimetry controls */
> > > > > > > >   	/* Keep the order of the 'case's the same as in
> > > > > > > >   	v4l2-controls.h!
> > > > > > > 
> > > > > > > */
> > > > > > > 
> > > > > > > > @@ -1518,6 +1519,10 @@ void v4l2_ctrl_fill(u32 id, const char
> > > > > > > > **name,
> > > > 
> > > > enum
> > > > 
> > > > > > > v4l2_ctrl_type *type,
> > > > > > > 
> > > > > > > >   	case V4L2_CID_STATELESS_HEVC_DECODE_PARAMS:
> > > > > > > >   		*type = 
V4L2_CTRL_TYPE_HEVC_DECODE_PARAMS;
> > > > > > > >   		break;
> > > > > > > > 
> > > > > > > > +	case V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS:
> > > > > > > > +		*type = V4L2_CTRL_TYPE_U32;
> > > > > > > > +		*flags |= V4L2_CTRL_FLAG_DYNAMIC_ARRAY;
> > > > > > > > +		break;
> > > > > > > > 
> > > > > > > >   	case V4L2_CID_STATELESS_VP9_COMPRESSED_HDR:
> > > > > > > >   		*type = 
V4L2_CTRL_TYPE_VP9_COMPRESSED_HDR;
> > > > > > > >   		break;
> > > > > > > > 
> > > > > > > > diff --git a/include/media/hevc-ctrls.h
> > > > > > > > b/include/media/hevc-ctrls.h
> > > > > > > > index a3c829ef531a..1319cb99ae3f 100644
> > > > > > > > --- a/include/media/hevc-ctrls.h
> > > > > > > > +++ b/include/media/hevc-ctrls.h
> > > > > > > > @@ -20,6 +20,7 @@
> > > > > > > > 
> > > > > > > >   #define V4L2_CID_STATELESS_HEVC_DECODE_PARAMS
> > 
> > (V4L2_CID_CODEC_BASE
> > 
> > > > > > > + 1012)
> > > > > > > 
> > > > > > > >   #define V4L2_CID_STATELESS_HEVC_DECODE_MODE
> > 
> > (V4L2_CID_CODEC_BASE
> > 
> > > > + 1015)
> > > > 
> > > > > > > >   #define V4L2_CID_STATELESS_HEVC_START_CODE
> > 
> > (V4L2_CID_CODEC_BASE + 1016)
> > 
> > > > > > > > +#define V4L2_CID_STATELESS_HEVC_ENTRY_POINT_OFFSETS
> > 
> > (V4L2_CID_CODEC_BASE
> > 
> > > > +
> > > > 
> > > > > > > 1017)
> > > > > > > 
> > > > > > > >   /* enum v4l2_ctrl_type type values */
> > > > > > > >   #define V4L2_CTRL_TYPE_HEVC_SPS 0x0120
> > > > > > > > 
> > > > > > > > @@ -318,6 +319,8 @@ struct v4l2_hevc_pred_weight_table {
> > > > > > > > 
> > > > > > > >    *
> > > > > > > >    * @bit_size: size (in bits) of the current slice data
> > > > > > > >    * @data_bit_offset: offset (in bits) to the video data in
> > > > > > > >    the current
> > > > 
> > > > slice
> > > > 
> > > > > > > data
> > > > > > > 
> > > > > > > > + * @num_entry_point_offsets: specifies the number of entry
> > > > > > > > point offset
> > > > 
> > > > syntax
> > > > 
> > > > > > > > + *			     elements in the slice 
header.
> > > > > > > > 
> > > > > > > >    * @nal_unit_type: specifies the coding type of the slice
> > > > > > > >    (B, P or I)
> > > > > > > >    * @nuh_temporal_id_plus1: minus 1 specifies a temporal
> > > > > > > >    identifier for
> > > > 
> > > > the
> > > > 
> > > > > > > NAL unit
> > > > > > > 
> > > > > > > >    * @slice_type: see V4L2_HEVC_SLICE_TYPE_{}
> > > > > > > > 
> > > > > > > > @@ -360,7 +363,7 @@ struct v4l2_hevc_pred_weight_table {
> > > > > > > > 
> > > > > > > >   struct v4l2_ctrl_hevc_slice_params {
> > > > > > > >   
> > > > > > > >   	__u32	bit_size;
> > > > > > > >   	__u32	data_bit_offset;
> > > > > > > > 
> > > > > > > > -
> > > > > > > > +	__u32	num_entry_point_offsets;
> > > > > > > > 
> > > > > > > >   	/* ISO/IEC 23008-2, ITU-T Rec. H.265: NAL unit 
header */
> > > > > > > >   	__u8	nal_unit_type;
> > > > > > > >   	__u8	nuh_temporal_id_plus1;

Powered by blists - more mailing lists