Re: [PATCH v2] [media] mtk-mdp: Fix g_/s_selection capture/compose logic

From: Hans Verkuil
Date: Mon Jun 19 2017 - 06:11:18 EST

On 06/19/2017 12:03 PM, Minghsiu Tsai wrote:
Hi, Hans,

On Fri, 2017-06-16 at 12:42 +0200, Hans Verkuil wrote:
On 05/12/17 04:42, Minghsiu Tsai wrote:
From: Daniel Kurtz <djkurtz@xxxxxxxxxxxx>

Experiments show that the:
(1) mtk-mdp uses the _MPLANE form of CAPTURE/OUTPUT

Please drop this, since this no longer applies to this patch.

I will remove it next version

(2) CAPTURE types use CROP targets, and OUTPUT types use COMPOSE targets

Are you really certain about this?

For m2m devices the output (i.e. memory to hardware) typically crops from memory
and the capture side (hardware to memory) composes into memory.

I.e.: for the output side you crop the part of the memory buffer that you want
to process and on the capture side you compose the result into a memory buffer:
i.e. the memory buffer might be 1920x1080, but you compose the decoder output
into a rectangle of 640x480 at offset 128x128 within that buffer (just an example).

CAPTURE using crop would be if, before the data is DMAed, the hardware decoder
output is cropped. E.g. if the stream fed to the decoder is 1920x1080, but you
want to only DMA a subselection of that, then that would be cropping, and it
would go to a memory buffer of the size of the crop selection.

OUTPUT using compose is highly unlikely: that means that the frame you give
is composed in a larger internal buffer with generated border data around it.
Very rare and really only something that a compositor of some sort would do.

That's strange. In v4l2-ioctl.c, v4l_g_crop()

The old crop API can't be used with most (all?) codec drivers. Codec drivers do
exactly the opposite of what the old crop API did.

It is the reason the selection API was created.

The old crop API was written for SDTV receivers where the hardware would crop
the received video. For output (rarely used) the hardware would compose the
image into a larger (PAL or NTSC sizes) frame.

Applications have to use the selection API to work with codec drivers. Just
ignore the old crop ioctls, they are not suitable for such hardware.

static int v4l_g_crop(const struct v4l2_ioctl_ops *ops,
struct file *file, void *fh, void *arg)
/* crop means compose for output devices */

ret = ops->vidioc_g_selection(file, fh, &s);

What exactly does the hardware do? Both for the encoder and for the decoder
case. Perhaps if I knew exactly what that is, then I can advise.

NV12M/YUV420M/MT21 -> MDP -> NV12M/YUV420M

The usage would be like this:
For decoder:
decoder -> MT21 -> MDP -> NV12M/YUV420M

For encoder:
NV12M/YUV420M -> MDP -> NV12M/YUV420M -> encoder

That doesn't help. Where in these two chains does the cropping (encoder) or
composing (decoder) take place? I assume it is the DMA engine that read from
a rectangle in the source buffer (encoder) or writes to a rectangle in the
destination buffer (decoder).

And in that case the current driver is correct.

I wonder if the g/s_crop description in the V4L2 Specification should get a
big fat warning that it doesn't apply to codec drivers and that the selection
API should be used instead. A patch for that is welcome!