17. Intel Image Processing Unit 3 (IPU3) Imaging Unit (ImgU) driver¶
Copyright © 2018 Intel Corporation
This file documents the Intel IPU3 (3rd generation Image Processing Unit) Imaging Unit drivers located under drivers/media/pci/intel/ipu3 (CIO2) as well as under drivers/staging/media/ipu3 (ImgU).
The Intel IPU3 found in certain Kaby Lake (as well as certain Sky Lake) platforms (U/Y processor lines) is made up of two parts namely the Imaging Unit (ImgU) and the CIO2 device (MIPI CSI2 receiver).
The CIO2 device receives the raw Bayer data from the sensors and outputs the frames in a format that is specific to the IPU3 (for consumption by the IPU3 ImgU). The CIO2 driver is available as drivers/media/pci/intel/ipu3/ipu3-cio2* and is enabled through the CONFIG_VIDEO_IPU3_CIO2 config option.
The Imaging Unit (ImgU) is responsible for processing images captured by the IPU3 CIO2 device. The ImgU driver sources can be found under drivers/staging/media/ipu3 directory. The driver is enabled through the CONFIG_VIDEO_IPU3_IMGU config option.
The two driver modules are named ipu3_csi2 and ipu3_imgu, respectively.
The drivers has been tested on Kaby Lake platforms (U/Y processor lines).
Both of the drivers implement V4L2, Media Controller and V4L2 sub-device interfaces. The IPU3 CIO2 driver supports camera sensors connected to the CIO2 MIPI CSI-2 interfaces through V4L2 sub-device sensor drivers.
The CIO2 is represented as a single V4L2 subdev, which provides a V4L2 subdev interface to the user space. There is a video node for each CSI-2 receiver, with a single media controller interface for the entire device.
The CIO2 contains four independent capture channel, each with its own MIPI CSI-2 receiver and DMA engine. Each channel is modelled as a V4L2 sub-device exposed to userspace as a V4L2 sub-device node and has two pads:
|0||sink||MIPI CSI-2 input, connected to the sensor subdev|
|1||source||Raw video capture, connected to the V4L2 video interface|
The V4L2 video interfaces model the DMA engines. They are exposed to userspace as V4L2 video device nodes.
17.2.1. Capturing frames in raw Bayer format¶
CIO2 MIPI CSI2 receiver is used to capture frames (in packed raw Bayer format) from the raw sensors connected to the CSI2 ports. The captured frames are used as input to the ImgU driver.
-- The IPU3 CSI2 receiver outputs the captured frames from the sensor in packed raw Bayer format that is specific to IPU3.
-- Multiple video nodes have to be operated simultaneously.
Let us take the example of ov5670 sensor connected to CSI2 port 0, for a 2592x1944 image capture.
Using the media contorller APIs, the ov5670 sensor is configured to send frames in packed raw Bayer format to IPU3 CSI2 receiver.
# This example assumes /dev/media0 as the CIO2 media device
# and that ov5670 sensor is connected to i2c bus 10 with address 0x36
export SDEV=$(media-ctl -d $MDEV -e "ov5670 10-0036")
# Establish the link for the media devices using media-ctl  media-ctl -d $MDEV -l "ov5670:0 -> ipu3-csi2 0:0"
# Set the format for the media devices media-ctl -d $MDEV -V "ov5670:0 [fmt:SGRBG10/2592x1944]"
media-ctl -d $MDEV -V "ipu3-csi2 0:0 [fmt:SGRBG10/2592x1944]"
media-ctl -d $MDEV -V "ipu3-csi2 0:1 [fmt:SGRBG10/2592x1944]"
Once the media pipeline is configured, desired sensor specific settings (such as exposure and gain settings) can be set, using the yavta tool.
yavta -w 0x009e0903 444 $SDEV
yavta -w 0x009e0913 1024 $SDEV
yavta -w 0x009e0911 2046 $SDEV
Once the desired sensor settings are set, frame captures can be done as below.
- yavta --data-prefix -u -c10 -n5 -I -s2592x1944 --file=/tmp/frame-#.bin
- -f IPU3_SGRBG10 $(media-ctl -d $MDEV -e "ipu3-cio2 0")
With the above command, 10 frames are captured at 2592x1944 resolution, with sGRBG10 format and output as IPU3_SGRBG10 format.
The captured frames are available as /tmp/frame-#.bin files.
The ImgU is represented as two V4L2 subdevs, each of which provides a V4L2 subdev interface to the user space.
Each V4L2 subdev represents a pipe, which can support a maximum of 2 streams. This helps to support advanced camera features like Continuous View Finder (CVF) and Snapshot During Video(SDV).
The ImgU contains two independent pipes, each modelled as a V4L2 sub-device exposed to userspace as a V4L2 sub-device node.
Each pipe has two sink pads and three source pads for the following purpose:
|0||sink||Input raw video stream|
|2||source||Output processed video stream|
|3||source||Output viewfinder video stream|
Each pad is connected to a corresponding V4L2 video interface, exposed to userspace as a V4L2 video device node.
17.3.1. Device operation¶
With ImgU, once the input video node ("ipu3-imgu 0/1":0, in <entity>:<pad-number> format) is queued with buffer (in packed raw Bayer format), ImgU starts processing the buffer and produces the video output in YUV format and statistics output on respective output nodes. The driver is expected to have buffers ready for all of parameter, output and statistics nodes, when input video node is queued with buffer.
At a minimum, all of input, main output, 3A statistics and viewfinder video nodes should be enabled for IPU3 to start image processing.
Each ImgU V4L2 subdev has the following set of video nodes.
17.3.2. input, output and viewfinder video nodes¶
The frames (in packed raw Bayer format specific to the IPU3) received by the input video node is processed by the IPU3 Imaging Unit and are output to 2 video nodes, with each targeting a different purpose (main output and viewfinder output).
Details onand the Bayer format specific to the IPU3 can be found in V4L2_PIX_FMT_IPU3_SBGGR10 ('ip3b'), V4L2_PIX_FMT_IPU3_SGBRG10 ('ip3g'), V4L2_PIX_FMT_IPU3_SGRBG10 ('ip3G'), V4L2_PIX_FMT_IPU3_SRGGB10 ('ip3r').
The driver supports V4L2 Video Capture Interface as defined at Interfaces.
Only the multi-planar API is supported. More details can be found at Single- and multi-planar APIs.
17.3.3. Parameters video node¶
The parameters video node receives the ImgU algorithm parameters that are used to configure how the ImgU algorithms process the image.
Details on processing parameters specific to the IPU3 can be found in V4L2_META_FMT_IPU3_PARAMS ('ip3p'), V4L2_META_FMT_IPU3_3A ('ip3s').
17.3.4. 3A statistics video node¶
3A statistics video node is used by the ImgU driver to output the 3A (auto focus, auto exposure and auto white balance) statistics for the frames that are being processed by the ImgU to user space applications. User space applications can use this statistics data to compute the desired algorithm parameters for the ImgU.
17.4. Configuring the Intel IPU3¶
The IPU3 ImgU pipelines can be configured using the Media Controller, defined at Part IV - Media Controller API.
17.4.1. Firmware binary selection¶
The firmware binary is selected using the V4L2_CID_INTEL_IPU3_MODE, currently defined in drivers/staging/media/ipu3/include/intel-ipu3.h . "VIDEO" and "STILL" modes are available.
17.4.2. Processing the image in raw Bayer format¶
184.108.40.206. Configuring ImgU V4L2 subdev for image processing¶
The ImgU V4L2 subdevs have to be configured with media controller APIs to have all the video nodes setup correctly.
Let us take "ipu3-imgu 0" subdev as an example.
media-ctl -d $MDEV -r
media-ctl -d $MDEV -l "ipu3-imgu 0 input":0 -> "ipu3-imgu 0":0
media-ctl -d $MDEV -l "ipu3-imgu 0":2 -> "ipu3-imgu 0 output":0
media-ctl -d $MDEV -l "ipu3-imgu 0":3 -> "ipu3-imgu 0 viewfinder":0
media-ctl -d $MDEV -l "ipu3-imgu 0":4 -> "ipu3-imgu 0 3a stat":0
Also the pipe mode of the corresponding V4L2 subdev should be set as desired (e.g 0 for video mode or 1 for still mode) through the control id 0x009819a1 as below.
yavta -w "0x009819A1 1" /dev/v4l-subdev7
RAW Bayer frames go through the following ImgU pipeline HW blocks to have the processed image output to the DDR memory.
RAW Bayer frame -> Input Feeder -> Bayer Down Scaling (BDS) -> Geometric Distortion Correction (GDC) -> DDR
The ImgU V4L2 subdev has to be configured with the supported resolutions in all the above HW blocks, for a given input resolution.
For a given supported resolution for an input frame, the Input Feeder, Bayer Down Scaling and GDC blocks should be configured with the supported resolutions. This information can be obtained by looking at the following IPU3 ImgU configuration table.
Under baseboard-poppy/media-libs/cros-camera-hal-configs-poppy/files/gcss directory, graph_settings_ov5670.xml can be used as an example.
The following steps prepare the ImgU pipeline for the image processing.
1. The ImgU V4L2 subdev data format should be set by using the VIDIOC_SUBDEV_S_FMT on pad 0, using the GDC width and height obtained above.
2. The ImgU V4L2 subdev cropping should be set by using the VIDIOC_SUBDEV_S_SELECTION on pad 0, with V4L2_SEL_TGT_CROP as the target, using the input feeder height and width.
3. The ImgU V4L2 subdev composing should be set by using the VIDIOC_SUBDEV_S_SELECTION on pad 0, with V4L2_SEL_TGT_COMPOSE as the target, using the BDS height and width.
For the ov5670 example, for an input frame with a resolution of 2592x1944 (which is input to the ImgU subdev pad 0), the corresponding resolutions for input feeder, BDS and GDC are 2592x1944, 2592x1944 and 2560x1920 respectively.
Once this is done, the received raw Bayer frames can be input to the ImgU V4L2 subdev as below, using the open source application v4l2n .
For an image captured with 2592x1944  resolution, with desired output resolution as 2560x1920 and viewfinder resolution as 2560x1920, the following v4l2n command can be used. This helps process the raw Bayer frames and produces the desired results for the main output image and the viewfinder output, in NV12 format.
v4l2n --pipe=4 --load=/tmp/frame-#.bin --open=/dev/video4 --fmt=type:VIDEO_OUTPUT_MPLANE,width=2592,height=1944,pixelformat=0X47337069 --reqbufs=type:VIDEO_OUTPUT_MPLANE,count:1 --pipe=1 --output=/tmp/frames.out --open=/dev/video5 --fmt=type:VIDEO_CAPTURE_MPLANE,width=2560,height=1920,pixelformat=NV12 --reqbufs=type:VIDEO_CAPTURE_MPLANE,count:1 --pipe=2 --output=/tmp/frames.vf --open=/dev/video6 --fmt=type:VIDEO_CAPTURE_MPLANE,width=2560,height=1920,pixelformat=NV12 --reqbufs=type:VIDEO_CAPTURE_MPLANE,count:1 --pipe=3 --open=/dev/video7 --output=/tmp/frames.3A --fmt=type:META_CAPTURE,? --reqbufs=count:1,type:META_CAPTURE --pipe=1,2,3,4 --stream=5
where /dev/video4, /dev/video5, /dev/video6 and /dev/video7 devices point to input, output, viewfinder and 3A statistics video nodes respectively.
17.4.3. Converting the raw Bayer image into YUV domain¶
The processed images after the above step, can be converted to YUV domain as below.
220.127.116.11. Main output frames¶
raw2pnm -x2560 -y1920 -fNV12 /tmp/frames.out /tmp/frames.out.ppm
where 2560x1920 is output resolution, NV12 is the video format, followed by input frame and output PNM file.
18.104.22.168. Viewfinder output frames¶
raw2pnm -x2560 -y1920 -fNV12 /tmp/frames.vf /tmp/frames.vf.ppm
where 2560x1920 is output resolution, NV12 is the video format, followed by input frame and output PNM file.
17.5. Example user space code for IPU3¶
User space code that configures and uses IPU3 is available here.
The source can be located under hal/intel directory.
17.6. Overview of IPU3 pipeline¶
IPU3 pipeline has a number of image processing stages, each of which takes a set of parameters as input. The major stages of pipelines are shown here:
The table below presents a description of the above algorithms.
|Optical Black Correction||Optical Black Correction block subtracts a pre-defined
value from the respective pixel values to obtain better
|Linearization||This algo block uses linearization parameters to
address non-linearity sensor effects. The Lookup table
table is defined in
|SHD||Lens shading correction is used to correct spatial
non-uniformity of the pixel response due to optical
lens shading. This is done by applying a different gain
for each pixel. The gain, black level etc are
|BNR||Bayer noise reduction block removes image noise by
applying a bilateral filter.
|ANR||Advanced Noise Reduction is a block based algorithm
that performs noise reduction in the Bayer domain. The
convolution matrix etc can be found in
|DM||Demosaicing converts raw sensor data in Bayer format
into RGB (Red, Green, Blue) presentation. Then add
outputs of estimation of Y channel for following stream
processing by Firmware. The struct is defined as
|Color Correction||Color Correction algo transforms sensor specific color
space to the standard "sRGB" color space. This is done
by applying 3x3 matrix defined in
|Gamma correction||Gamma correction
|CSC||Color space conversion transforms each pixel from the
RGB primary presentation to YUV (Y: brightness,
UV: Luminance) presentation. This is done by applying
a 3x3 matrix defined in
|CDS||Chroma down sampling
After the CSC is performed, the Chroma Down Sampling
is applied for a UV plane down sampling by a factor
of 2 in each direction for YUV 4:2:0 using a 4x2
|CHNR||Chroma noise reduction
This block processes only the chrominance pixels and
performs noise reduction by cleaning the high
|TCC||Total color correction as defined in struct
|XNR3||eXtreme Noise Reduction V3 is the third revision of
noise reduction algorithm used to improve image
quality. This removes the low frequency noise in the
captured image. Two related structs are being defined,
|TNR||Temporal Noise Reduction block compares successive
frames in time to remove anomalies / noise in pixel
Other often encountered acronyms not listed in above table:
- Accelerator cluster
- Auto white balance filter response statistics
- Bayer downscaler parameters
- Color correction matrix coefficients
- Image enhancement filter directed
- Optical black level compensation
- Output system configuration
- Region of interest
- Y down sampling
- Y-tone mapping
A few stages of the pipeline will be executed by firmware running on the ISP processor, while many others will use a set of fixed hardware blocks also called accelerator cluster (ACC) to crunch pixel data and produce statistics.
ACC parameters of individual algorithms, as defined by
ipu3_uapi_acc_param, can be chosen to be applied by the user
space through struct
ipu3_uapi_flags embedded in
ipu3_uapi_params structure. For parameters that are configured as
not enabled by the user space, the corresponding structs are ignored by the
driver, in which case the existing configuration of the algorithm will be