On 17/10/2022 14:06, Kieran Bingham wrote:
Quoting Bryan O'Donoghue (2022-10-17 01:16:05)
On 13/10/2022 13:12, quic_mmitkov@xxxxxxxxxxx wrote:
From: Milen Mitkov <quic_mmitkov@xxxxxxxxxxx>
For v4:
- fixes the warning reported by the kernel test robot
- tiny code change to enable the vc functionality with the partially-applied
multistream patches on linux-next (tested on tag:next-20221010)
For v3:
- setting the sink pad format on the CSID entity will now propagate the
format to the source pads to keep the subdev in a valid internal state.
- code syntax improvements
For v2:
- code syntax improvements
- The info print for the enabled VCs was demoted to a dbg print. Can be
enabled with dynamic debug, e.g.:
echo "file drivers/media/platform/qcom/camss/* +p" > /sys/kernel/debug/dynamic_debug/control
NOTE: These changes depend on the multistream series, that as of yet
is still not merged upstream. However, part of the
multistream patches are merged in linux-next (tested on
tag:next-20221010), including the patch that introduces the
video_device_pipeline_alloc_start() functionality. This allows
applying and using this series on linux-next without applying the
complete multistream set.
The CSID hardware on SM8250 can demux the input data stream into
maximum of 4 multiple streams depending on virtual channel (vc)
or data type (dt) configuration.
Situations in which demuxing is useful:
- HDR sensors that produce a 2-frame HDR output, e.g. a light and a dark frame
(the setup we used for testing, with the imx412 sensor),
or a 3-frame HDR output - light, medium-lit, dark frame.
- sensors with additional metadata that is streamed over a different
virtual channel/datatype.
- sensors that produce frames with multiple resolutions in the same pixel
data stream
With these changes, the CSID entity has, as it did previously, a single
sink port (0), and always exposes 4 source ports (1, 2,3, 4). The
virtual channel configuration is determined by which of the source ports
are linked to an output VFE line. For example, the link below will
configure the CSID driver to enable vc 0 and vc 1:
media-ctl -l '"msm_csid0":1->"msm_vfe0_rdi0":0[1]'
media-ctl -l '"msm_csid0":2->"msm_vfe0_rdi1":0[1]'
which will be demuxed and propagated into /dev/video0
and /dev/video1 respectively. With this, the userspace can use
any normal V4L2 client app to start/stop/queue/dequeue from these
video nodes. Tested with the yavta app.
The format of each RDI channel of the used VFE(s) (msm_vfe0_rdi0,
msm_vfe0_rdi1,...) must also be configured explicitly.
Note that in order to keep a valid internal subdevice state,
setting the sink pad format of the CSID subdevice will propagate
this format to the source pads. However, since the CSID hardware
can demux the input stream into several streams each of which can
be a different format, in that case each source pad's
format must be set individually, e.g.:
media-ctl -V '"msm_csid0":1[fmt:SRGGB10/3840x2160]'
media-ctl -V '"msm_csid0":2[fmt:SRGGB10/960x540]'
Milen Mitkov (4):
media: camss: sm8250: Virtual channels for CSID
media: camss: vfe: Reserve VFE lines on stream start and link to CSID
media: camss: vfe-480: Multiple outputs support for SM8250
media: camss: sm8250: Pipeline starting and stopping for multiple
virtual channels
.../platform/qcom/camss/camss-csid-gen2.c | 54 ++++++++++------
.../media/platform/qcom/camss/camss-csid.c | 44 +++++++++----
.../media/platform/qcom/camss/camss-csid.h | 11 +++-
.../media/platform/qcom/camss/camss-vfe-480.c | 61 ++++++++++++-------
drivers/media/platform/qcom/camss/camss-vfe.c | 7 +++
.../media/platform/qcom/camss/camss-video.c | 21 ++++++-
drivers/media/platform/qcom/camss/camss.c | 2 +-
7 files changed, 140 insertions(+), 60 deletions(-)
Hi Milen
The set applies to next-20221013 including patch#4.
I can confirm it doesn't break anything for me - though my sensor is a
bog-standard imx577 so it doesn't have any VC support.
Interesting though - the IMX477 has the ability to convey metadata on a
separate VC... That's actually the thing holding back the RPi IMX477
driver from mainline, as the way it was anticipated to support multiple
data streams is with the new multiplexed streams API.
And I think we inferred that the IMX577 and IMX477 are closely related,
so should have similar capabilities for obtaining metadata channels?
Hmm I was not aware of that.
If we could import the rpi/imx477.c code into upstrea/imx412.c it might
be possible
The core init is very similar
https://github.com/raspberrypi/linux/blob/rpi-5.19.y/drivers/media/i2c/imx477.c#L167
https://github.com/raspberrypi/linux/blob/rpi-5.19.y/drivers/media/i2c/imx412.c#L160
Maybe it would be possible to apply the rest of the imx477 config on-top
as a POC
https://github.com/raspberrypi/linux/blob/rpi-5.19.y/drivers/media/i2c/imx477.c#L479
The similary is born out by the shared init code I can see in Leopard
imaging's driver, I'm not sure if it supports virtual-channels - I'll
have a look, though.
What's in the imx477 meta-data ?
@Milen if you have the imx577 datasheet - I don't - perhaps we could
cherry-pick some of the code from imx477 and get the imx412.c->imx577
dumping VC data out with the RB5 camera mezzanine.
---
bod