Re: [PATCH] i.MX6: Support 16-bit BT.1120 video input

[Philipp Zabel <p.zabel@pengutronix.de>]

Hi Krzysztof,

On Wed, 2021-10-06 at 12:52 +0200, Krzysztof Hałasa wrote:
> Hi Philipp,
> 
> Philipp Zabel <p.zabel@pengutronix.de> writes:
> 
> > > + * - BT.656 and BT.1120 (8/10-bit YUV422) data can always be processed
> > > + *   on-the-fly (converted to YUV420)
> > 
> > This comment seems misleading. The CSI converts to YUV 4:4:4 internally.
> 
> Well... this is surprising. You mean "on the internal bus", don't you?

Yes, anything apart from the bayer/generic data modes, internally
everything is converted into 32-bit YUVA/RGBA pixels (according to the
Reference Manual, 37.4.2.3 FCW & FCR). That is represented by the
MEDIA_BUS_FMT_AYUV8_1X32 media bus format at the CSI source pads.

> Please correct me if the following is wrong:
> 
> I always though that the "on-the-fly processing", in case of YUV422,
> means in practice I can get YUV420 out of the IPU, without a need to do
> e.g. NEON conversion.

That is done in the IDMAC, which can write any supported YUV format from
the internal YUV pixels (if not in bayer/generic data mode).

> I know I can get the original YUV422 as well,
> using the "generic data" mode, but it's incompatible with the CODA H.264
> encoder.

You should also be able to store the YUV formatted pixels as NV12, NV16,
YUYV, etc.

> Ok, the DQRM (37.4.3.2.1) states that for parallel YUV the output from
> CSI is always YUV444.

Ack.

> Then 37.4.3.9 says that the only YUV422 way is to use 16-bit "generic
> data". This doesn't seem to be very true, however I'm not exactly sure
> about the "on-the-fly" thing.

I think that statement is limited to the parallel 16-bit interface in
hsync/vsync mode, whereas in bt.656 / bt.1120 mode the interface
operates as if the two components were clocked in as two separate 8-bit
(or 10-bit) values.

> The fact is the patch works.
> Also, the CSIx_SENS_DATA_FORMAT field in IPUx_CSIx_SENS_CONF register
> shows YUV422 YUYV and UYVY input data formats, clearly separate from
> "Bayer of Generic data".
> 
> DQIEC, 4.12.10.1, isn't very clear either:
> 8) YCbCr 20-bit (10-bit Y + 10-bit U/V) is supported with BT.1120 only
> 7) YCbCr 16-bit is supported under the same conditions as 8)
> 6) YCbCr 16-bit (= YUV422) is also supported as "generic-data"
>    (no on-the-fly processing). This seems to imply 8) and 7) are
>    supported WITH o-t-f-p (and obviously I have tested it, 16-bit only).
>
> I think I will just remove the comment :-)

That sounds good to me.

regards
Philipp