Message-ID: <20180814083054.GB21634@phenom.ffwll.local>
Date:   Tue, 14 Aug 2018 10:30:54 +0200
From:   Daniel Vetter <daniel@...ll.ch>
To:     Brian Starkey <brian.starkey@....com>
Cc:     dri-devel@...ts.freedesktop.org, alexandru-cosmin.gheorghe@....com,
        ayan.halder@....com, james.qian.wang@....com, daniel@...ishbar.org,
        daniel@...ll.ch, krh@...planet.net,
        Gustavo Padovan <gustavo@...ovan.org>,
        Maarten Lankhorst <maarten.lankhorst@...ux.intel.com>,
        Sean Paul <seanpaul@...omium.org>,
        David Airlie <airlied@...ux.ie>, linux-kernel@...r.kernel.org,
        raymond.smith@....com, david.garbett@....com, lisa.wu@....com,
        matt.szczesiak@....com, charles.xu@....com
Subject: Re: [RFC PATCH] drm/fourcc: Add remaining fourccs for Mali

On Fri, Aug 10, 2018 at 04:00:17PM +0100, Brian Starkey wrote:
> This RFC adds the remaining fourcc codes which are needed to cover all
> currently supported formats in Mali IPs.
> 
> There was quite a lengthy discussion on IRC about it - please treat this
> as RFC and read the commentary below!
> 
> This patch is on-top of
> https://lists.freedesktop.org/archives/dri-devel/2018-July/184598.html,
> but note I have changed the fourcc code for XVYU2101010 to be more
> consistent. We can of course update the previous patch after we have
> resolved this discussion.
> 
> ---
> 
> As we look to enable AFBC using DRM format modifiers, we run into
> problems which we've historically handled via vendor-private details
> (i.e. gralloc, on Android).
> 
> AFBC (as an encoding) is fully flexible, and for example YUV data can
> be encoded into 1, 2 or 3 encoded "planes", much like the linear
> equivalents. Component order is also meaningful, as AFBC doesn't
> necessarily care about what each "channel" of the data it encodes
> contains. Therefore ABGR8888 and RGBA8888 can be encoded in AFBC with
> different representations. Similarly, 'X' components may be encoded
> into AFBC streams in cases where a decoder expects to decode a 4th
> component.
> 
> In addition, AFBC is a licensable IP, meaning that to support the
> ecosystem we need to ensure that _all_ AFBC users are able to describe
> the encodings that they need. This is much better achieved by
> preserving meaning in the fourcc codes when they are combined with an
> AFBC modifier.
> 
> In essence, we want to use the modifier to describe the parameters of
> the AFBC encode/decode, and use the fourcc code to describe the data
> being encoded/decoded.
> 
> To do anything different would be to introduce redundancy - we would
> need to duplicate in the modifier information which is _already_
> conveyed clearly and non-ambigiously by a fourcc code.
> 
> I hope that for RGB this is non-controversial.
> (BGRA8888 + MODIFIER_AFBC) is a different format from
> (RGBA8888 + MODIFIER_AFBC).
> 
> Possibly more controversial is that (XBGR8888 + MODIFIER_AFBC)
> is different from (BGR888 + MODIFIER_AFBC). I understand that in some
> schemes it is not the case - but in AFBC it is so.
> 
> Where we run into problems is where there are not already fourcc codes
> which represent the data which the AFBC encoder/decoder is processing.
> To that end, we want to introduce new fourcc codes to describe the
> data being encoded/decoded, in the places where none of the existing
> fourcc codes are applicable.
> 
> Where we don't support an equivalent non-compressed layout, or where
> no "obvious" linear layout exists, we are proposing adding fourcc
> codes which have no associated linear layout - because any layout we
> proposed would be completely arbitrary.
> 
> Some formats are following the naming conventions from [2].
> 
> The summary of the new formats is:
>  DRM_FORMAT_VUY888 - Packed 8-bit YUV 4:4:4. Y followed by U then V.
>  DRM_FORMAT_VUY101010 - Packed 10-bit YUV 4:4:4. Y followed by U then
>                         V. No defined linear encoding.
>  DRM_FORMAT_Y210 - Packed 10-bit YUV 4:2:2. Y followed by U (then Y)
>                    then V. 10-bit samples in 16-bit words.
>  DRM_FORMAT_P210 - Semi-planar 10-bit YUV 4:2:2. Y plane, followed by
>                    interleaved U-then-V plane. 10-bit samples in
>                    16-bit words.
>  DRM_FORMAT_YUV420_8BIT - Packed 8-bit YUV 4:2:0. Y followed by U then
>                           V. No defined linear encoding
>  DRM_FORMAT_YUV420_10BIT - Packed 10-bit YUV 4:2:0. Y followed by U
>                            then V. No defined linear encoding
>  DRM_FORMAT_P010 - Semi-planar 10-bit YUV 4:2:0. Y plane, followed by
>                    interleaved U-then-V plane. Added in [1].

Your use of 4:4:4 and friends for planar formats is very confusing in the
context of drm_fourcc.h, since that's how we spec out bit layout in the
drm_fourcc.h file. Please use YUV 444 sub-sampling or similar to
disambiguate. The patch itself isn't confusing like this.

> 
> Please also note that in the absence of AFBC, we would still need to
> add Y210, P210 and P010.
> 
> Full rationale follows:
> 
> YUV 4:4:4 8-bit, 1-plane - New format
> -------------------------------------
>  The currently defined AYUV format encodes a 4th alpha component,
>  which makes it unsuitable for representing a 3-component YUV 4:4:4
>  AFBC stream.
> 
>  The proposed[1] XYUV format which is supported by Mali-DP in linear
>  layout is also unsuitable, because the component order is the
>  opposite of the AFBC version, and it encodes a 4th 'X' component.
> 
>  DRM_FORMAT_VUY888 is the "obvious" format for a 3-component, packed,
>  YUV 444 8-bit format, with the component order which our HW expects to
>  encode/decode. It conforms to the same naming convention as the
>  existing packed YUV 4:4:4 format.
>  The naming here is meant to be consistent with DRM_FORMAT_AYUV and
>  DRM_FORMAT_XYUV[1]
> 
> YUV 4:4:4 10-bit, 1-plane - New format
> --------------------------------------
>  There is no currently-defined YUV 4:4:4 10-bit format in
>  drm_fourcc.h, irrespective of number of planes.
> 
>  The proposed[1] XVYU2101010 format which is supported by Mali-DP in
>  linear layout uses the wrong component order, and also encodes a 4th
>  'X' component, which doesn't match the AFBC version of YUV 4:4:4
>  10-bit which we support.
> 
>  There is no "obvious" linear encoding for a 3-component 10:10:10
>  packed format, and so DRM_FORMAT_VUY101010 defines a component
>  order, but not a bit encoding. Again, the naming is meant to be
>  consistent with DRM_FORMAT_AYUV.
> 
> YUV 4:2:2 8-bit, 1-plane
> ------------------------
>  The existing DRM_FORMAT_YUYV (and the other component orders) are
>  single-planar YUV 4:2:2 8-bit formats. Following the convention of
>  the component orders of the RGB formats, YUYV has the correct
>  component order for our AFBC encoding (Y followed by U followed by
>  V). We can use YUYV for AFBC YUV 4:2:2 8-bit.
> 
> YUV 4:2:2 10-bit, 1-plane - New format
> --------------------------------------
>  There is no currently-defined YUV 4:2:2 10-bit format in drm_fourcc.h
> 
>  DRM_FORMAT_Y210 is analogous to YUYV, but with 10-bits per sample
>  packed into the upper 10-bits of 16-bit samples. This format is
>  supported in both linear and AFBC by Mali GPUs.
> 
> YUV 4:2:2 10-bit, 2-plane - New format
> --------------------------------------
>  There is no currently-defined YUV 4:2:2 10-bit format in drm_fourcc.h
> 
>  DRM_FORMAT_P210 is analogous to NV16, but with 10-bits per sample
>  packed into the upper 10-bits of 16-bit samples. This format is
>  supported in both Linear and AFBC by Mali GPUs.
> 
> YUV 4:2:0 8-bit, 1-plane - New format
> -------------------------------------
>  There is no currently defined single-planar YUV 4:2:0, 8-bit format
>  in drm_fourcc.h. There's differing opinions on whether using the
>  existing fourcc-implied n_planes where possible is a good idea or
>  not when using modifiers.
> 
>  For me, it's much more "obvious" to use NV12 for 2-plane AFBC and
>  YUV420 for 3-plane AFBC. This keeps the aforementioned separation
>  between the AFBC codec settings (in the modifier) and the pixel data
>  format (in the fourcc). With different vendors using AFBC, this helps
>  to ensure that there is no confusion in interoperation. It also
>  ensures that the AFBC modifiers describe AFBC itself (which is a
>  licensable component), and not implementation details which are not
>  defined by AFBC.
> 
>  The proposed[1] X0L0 format which Mali-DP supports with Linear layout
>  is unsuitable, as it contains a 4th 'X' component, and our AFBC
>  decoder expects only 3 components.
> 
>  To that end, we propose a new YUV 4:2:0 8-bit format. There is no
>  "obvious" linear encoding for a 3-component 8:8:8, 4:2:0, packed format,
>  and so DRM_FORMAT_YUV420_8BIT defines a component order, but not a
>  bit encoding. I'm happy to hear different naming suggestions.
> 
> YUV 4:2:0 8-bit, 2-, 3-plane
> ----------------------------
>  These already exist, we can use NV12 and YUV420.
> 
> YUV 4:2:0 10-bit, 1-plane - New format
> -------------------------------------
>  As above, no current definition exists, and X0L2 encodes a 4th 'X'
>  channel.
> 
>  Analogous to DRM_FORMAT_YUV420_8BIT, we define DRM_FORMAT_YUV420_10BIT.
> 
> YUV 4:2:0 10-bit, 2-plane - New format
> --------------------------------------
>  The proposed[1] (and relatively common) P010 is a 2-plane, 10-bit YUV
>  format, which has the correct component order for our AFBC encoding.
> 
> Thanks and best regards,
> -Brian
> 
> [1] https://lists.freedesktop.org/archives/dri-devel/2018-July/184598.html
> [2] https://docs.microsoft.com/en-us/windows/desktop/medfound/10-bit-and-16-bit-yuv-video-formats

I'm kinda missing the patch to add AFBC modifier itself, that one would be
interesting as patch 2 in this series.

For this one here I'd say fix up the confusion in the text, keep the
entire text as commit message, add sob, and you get my ack. Also, it might
be good to put the full text of how AFBC modifier is supposed to be used
in Documentation/gpu/drm-uapi.rst (or a new file/kernel-doc linked from
there). Just to make sure as many people as possible will stumble over it,
and hence increase the odds this won't end in total confusion. Which I
agree, is a big risk for AFBC due to the licensed ecosystem nature of the
beast here.
-Daniel

> 
>  include/uapi/drm/drm_fourcc.h | 15 ++++++++++++++-
>  1 file changed, 14 insertions(+), 1 deletion(-)
> 
> diff --git a/include/uapi/drm/drm_fourcc.h b/include/uapi/drm/drm_fourcc.h
> index 213357940424..a44bbb032171 100644
> --- a/include/uapi/drm/drm_fourcc.h
> +++ b/include/uapi/drm/drm_fourcc.h
> @@ -110,11 +110,14 @@ extern "C" {
>  #define DRM_FORMAT_YVYU		fourcc_code('Y', 'V', 'Y', 'U') /* [31:0] Cb0:Y1:Cr0:Y0 8:8:8:8 little endian */
>  #define DRM_FORMAT_UYVY		fourcc_code('U', 'Y', 'V', 'Y') /* [31:0] Y1:Cr0:Y0:Cb0 8:8:8:8 little endian */
>  #define DRM_FORMAT_VYUY		fourcc_code('V', 'Y', 'U', 'Y') /* [31:0] Y1:Cb0:Y0:Cr0 8:8:8:8 little endian */
> +#define DRM_FORMAT_Y210		fourcc_code('Y', '2', '1', '0') /* [63:0] Cr0:0:Y1:0:Cb0:0:Y0:0 10:6:10:6:10:6:10:6 little endian */
>  
>  /* packed YCbCr444 */
>  #define DRM_FORMAT_AYUV		fourcc_code('A', 'Y', 'U', 'V') /* [31:0] A:Y:Cb:Cr 8:8:8:8 little endian */
>  #define DRM_FORMAT_XYUV8888	fourcc_code('X', 'Y', 'U', 'V') /* [31:0] X:Y:Cb:Cr 8:8:8:8 little endian */
> -#define DRM_FORMAT_XVYU2101010	fourcc_code('V', 'U', '3', '0') /* [31:0] X:Cr:Y:Cb 2:10:10:10 little endian */
> +#define DRM_FORMAT_VUY888	fourcc_code('V', 'U', '2', '4') /* [23:0] Cr:Cb:Y 8:8:8 little endian */
> +#define DRM_FORMAT_XVYU2101010	fourcc_code('X', 'V', '3', '0') /* [31:0] X:Cr:Y:Cb 2:10:10:10 little endian */
> +#define DRM_FORMAT_VUY101010    fourcc_code('V', 'U', '3', '0') /* Y followed by U then V, 10:10:10. Non-linear modifier only */
>  
>  /*
>   * packed YCbCr420 2x2 tiled formats
> @@ -131,6 +134,15 @@ extern "C" {
>  /* [63:0]   X3:X2:Y3:Cr0:Y2:X1:X0:Y1:Cb0:Y0  1:1:10:10:10:1:1:10:10:10 little endian */
>  #define DRM_FORMAT_X0L2		fourcc_code('X', '0', 'L', '2')
>  
> +/*
> + * 1-plane YUV 4:2:0
> + * In these formats, the component ordering is specified (Y, followed by U
> + * then V), but the exact Linear layout is undefined.
> + * These formats can only be used with a non-Linear modifier.
> + */
> +#define DRM_FORMAT_YUV420_8BIT	fourcc_code('Y', 'U', '0', '8')
> +#define DRM_FORMAT_YUV420_10BIT	fourcc_code('Y', 'U', '1', '0')
> +
>  /*
>   * 2 plane RGB + A
>   * index 0 = RGB plane, same format as the corresponding non _A8 format has
> @@ -164,6 +176,7 @@ extern "C" {
>   * Y plane: [63:0] Y3:0:Y2:0:Y1:0:Y0:0, 10:6:10:6:10:6:10:6
>   * CrCb plane: [63:0] Cr2:0:Cb2:0:Cr0:0:Cb0:0, 10:6:10:6:10:6:10:6
>   */
> +#define DRM_FORMAT_P210		fourcc_code('P', '2', '1', '0') /* 2x1 subsampled Cr:Cb plane */
>  #define DRM_FORMAT_P010		fourcc_code('P', '0', '1', '0') /* 2x2 subsampled Cr:Cb plane */
>  
>  /*
> -- 
> 2.17.1
> 

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch

Powered by blists - more mailing lists