  <title>Sub-device Interface</title>

  <note>
    <title>Experimental</title>
    <para>This is an <link linkend="experimental">experimental</link>
    interface and may change in the future.</para>
  </note>

  <para>The complex nature of V4L2 devices, where hardware is often made of
  several integrated circuits that need to interact with each other in a
  controlled way, leads to complex V4L2 drivers. The drivers usually reflect
  the hardware model in software, and model the different hardware components
  as software blocks called sub-devices.</para>

  <para>V4L2 sub-devices are usually kernel-only objects. If the V4L2 driver
  implements the media device API, they will automatically inherit from media
  entities. Applications will be able to enumerate the sub-devices and discover
  the hardware topology using the media entities, pads and links enumeration
  API.</para>

  <para>In addition to make sub-devices discoverable, drivers can also choose
  to make them directly configurable by applications. When both the sub-device
  driver and the V4L2 device driver support this, sub-devices will feature a
  character device node on which ioctls can be called to
  <itemizedlist>
    <listitem><para>query, read and write sub-devices controls</para></listitem>
    <listitem><para>subscribe and unsubscribe to events and retrieve them</para></listitem>
    <listitem><para>negotiate image formats on individual pads</para></listitem>
  </itemizedlist>
  </para>

  <para>Sub-device character device nodes, conventionally named
  <filename>/dev/v4l-subdev*</filename>, use major number 81.</para>

  <section>
    <title>Controls</title>
    <para>Most V4L2 controls are implemented by sub-device hardware. Drivers
    usually merge all controls and expose them through video device nodes.
    Applications can control all sub-devices through a single interface.</para>

    <para>Complex devices sometimes implement the same control in different
    pieces of hardware. This situation is common in embedded platforms, where
    both sensors and image processing hardware implement identical functions,
    such as contrast adjustment, white balance or faulty pixels correction. As
    the V4L2 controls API doesn't support several identical controls in a single
    device, all but one of the identical controls are hidden.</para>

    <para>Applications can access those hidden controls through the sub-device
    node with the V4L2 control API described in <xref linkend="control" />. The
    ioctls behave identically as when issued on V4L2 device nodes, with the
    exception that they deal only with controls implemented in the sub-device.
    </para>

    <para>Depending on the driver, those controls might also be exposed through
    one (or several) V4L2 device nodes.</para>
  </section>

  <section>
    <title>Events</title>
    <para>V4L2 sub-devices can notify applications of events as described in
    <xref linkend="event" />. The API behaves identically as when used on V4L2
    device nodes, with the exception that it only deals with events generated by
    the sub-device. Depending on the driver, those events might also be reported
    on one (or several) V4L2 device nodes.</para>
  </section>

  <section id="pad-level-formats">
    <title>Pad-level Formats</title>

    <warning><para>Pad-level formats are only applicable to very complex device that
    need to expose low-level format configuration to user space. Generic V4L2
    applications do <emphasis>not</emphasis> need to use the API described in
    this section.</para></warning>

    <note><para>For the purpose of this section, the term
    <wordasword>format</wordasword> means the combination of media bus data
    format, frame width and frame height.</para></note>

    <para>Image formats are typically negotiated on video capture and
    output devices using the format and <link
    linkend="vidioc-subdev-g-selection">selection</link> ioctls. The
    driver is responsible for configuring every block in the video
    pipeline according to the requested format at the pipeline input
    and/or output.</para>

    <para>For complex devices, such as often found in embedded systems,
    identical image sizes at the output of a pipeline can be achieved using
    different hardware configurations. One such example is shown on
    <xref linkend="pipeline-scaling" />, where
    image scaling can be performed on both the video sensor and the host image
    processing hardware.</para>

    <figure id="pipeline-scaling">
      <title>Image Format Negotiation on Pipelines</title>
      <mediaobject>
	<imageobject>
	  <imagedata fileref="pipeline.pdf" format="PS" />
	</imageobject>
	<imageobject>
	  <imagedata fileref="pipeline.png" format="PNG" />
	</imageobject>
	<textobject>
	  <phrase>High quality and high speed pipeline configuration</phrase>
	</textobject>
      </mediaobject>
    </figure>

    <para>The sensor scaler is usually of less quality than the host scaler, but
    scaling on the sensor is required to achieve higher frame rates. Depending
    on the use case (quality vs. speed), the pipeline must be configured
    differently. Applications need to configure the formats at every point in
    the pipeline explicitly.</para>

    <para>Drivers that implement the <link linkend="media-controller-intro">media
    API</link> can expose pad-level image format configuration to applications.
    When they do, applications can use the &VIDIOC-SUBDEV-G-FMT; and
    &VIDIOC-SUBDEV-S-FMT; ioctls. to negotiate formats on a per-pad basis.</para>

    <para>Applications are responsible for configuring coherent parameters on
    the whole pipeline and making sure that connected pads have compatible
    formats. The pipeline is checked for formats mismatch at &VIDIOC-STREAMON;
    time, and an &EPIPE; is then returned if the configuration is
    invalid.</para>

    <para>Pad-level image format configuration support can be tested by calling
    the &VIDIOC-SUBDEV-G-FMT; ioctl on pad 0. If the driver returns an &EINVAL;
    pad-level format configuration is not supported by the sub-device.</para>

    <section>
      <title>Format Negotiation</title>

      <para>Acceptable formats on pads can (and usually do) depend on a number
      of external parameters, such as formats on other pads, active links, or
      even controls. Finding a combination of formats on all pads in a video
      pipeline, acceptable to both application and driver, can't rely on formats
      enumeration only. A format negotiation mechanism is required.</para>

      <para>Central to the format negotiation mechanism are the get/set format
      operations. When called with the <structfield>which</structfield> argument
      set to <constant>V4L2_SUBDEV_FORMAT_TRY</constant>, the
      &VIDIOC-SUBDEV-G-FMT; and &VIDIOC-SUBDEV-S-FMT; ioctls operate on a set of
      formats parameters that are not connected to the hardware configuration.
      Modifying those 'try' formats leaves the device state untouched (this
      applies to both the software state stored in the driver and the hardware
      state stored in the device itself).</para>

      <para>While not kept as part of the device state, try formats are stored
      in the sub-device file handles. A &VIDIOC-SUBDEV-G-FMT; call will return
      the last try format set <emphasis>on the same sub-device file
      handle</emphasis>. Several applications querying the same sub-device at
      the same time will thus not interact with each other.</para>

      <para>To find out whether a particular format is supported by the device,
      applications use the &VIDIOC-SUBDEV-S-FMT; ioctl. Drivers verify and, if
      needed, change the requested <structfield>format</structfield> based on
      device requirements and return the possibly modified value. Applications
      can then choose to try a different format or accept the returned value and
      continue.</para>

      <para>Formats returned by the driver during a negotiation iteration are
      guaranteed to be supported by the device. In particular, drivers guarantee
      that a returned format will not be further changed if passed to an
      &VIDIOC-SUBDEV-S-FMT; call as-is (as long as external parameters, such as
      formats on other pads or links' configuration are not changed).</para>

      <para>Drivers automatically propagate formats inside sub-devices. When a
      try or active format is set on a pad, corresponding formats on other pads
      of the same sub-device can be modified by the driver. Drivers are free to
      modify formats as required by the device. However, they should comply with
      the following rules when possible:
      <itemizedlist>
        <listitem><para>Formats should be propagated from sink pads to source pads.
	Modifying a format on a source pad should not modify the format on any
	sink pad.</para></listitem>
        <listitem><para>Sub-devices that scale frames using variable scaling factors
	should reset the scale factors to default values when sink pads formats
	are modified. If the 1:1 scaling ratio is supported, this means that
	source pads formats should be reset to the sink pads formats.</para></listitem>
      </itemizedlist>
      </para>

      <para>Formats are not propagated across links, as that would involve
      propagating them from one sub-device file handle to another. Applications
      must then take care to configure both ends of every link explicitly with
      compatible formats. Identical formats on the two ends of a link are
      guaranteed to be compatible. Drivers are free to accept different formats
      matching device requirements as being compatible.</para>

      <para><xref linkend="sample-pipeline-config" />
      shows a sample configuration sequence for the pipeline described in
      <xref linkend="pipeline-scaling" /> (table
      columns list entity names and pad numbers).</para>

      <table pgwide="0" frame="none" id="sample-pipeline-config">
	<title>Sample Pipeline Configuration</title>
	<tgroup cols="3">
	  <colspec colname="what"/>
	  <colspec colname="sensor-0" />
	  <colspec colname="frontend-0" />
	  <colspec colname="frontend-1" />
	  <colspec colname="scaler-0" />
	  <colspec colname="scaler-1" />
	  <thead>
	    <row>
	      <entry></entry>
	      <entry>Sensor/0</entry>
	      <entry>Frontend/0</entry>
	      <entry>Frontend/1</entry>
	      <entry>Scaler/0</entry>
	      <entry>Scaler/1</entry>
	    </row>
	  </thead>
	  <tbody valign="top">
	    <row>
	      <entry>Initial state</entry>
	      <entry>2048x1536</entry>
	      <entry>-</entry>
	      <entry>-</entry>
	      <entry>-</entry>
	      <entry>-</entry>
	    </row>
	    <row>
	      <entry>Configure frontend input</entry>
	      <entry>2048x1536</entry>
	      <entry><emphasis>2048x1536</emphasis></entry>
	      <entry><emphasis>2046x1534</emphasis></entry>
	      <entry>-</entry>
	      <entry>-</entry>
	    </row>
	    <row>
	      <entry>Configure scaler input</entry>
	      <entry>2048x1536</entry>
	      <entry>2048x1536</entry>
	      <entry>2046x1534</entry>
	      <entry><emphasis>2046x1534</emphasis></entry>
	      <entry><emphasis>2046x1534</emphasis></entry>
	    </row>
	    <row>
	      <entry>Configure scaler output</entry>
	      <entry>2048x1536</entry>
	      <entry>2048x1536</entry>
	      <entry>2046x1534</entry>
	      <entry>2046x1534</entry>
	      <entry><emphasis>1280x960</emphasis></entry>
	    </row>
	  </tbody>
	</tgroup>
      </table>

      <para>
      <orderedlist>
	<listitem><para>Initial state. The sensor output is set to its native 3MP
	resolution. Resolutions on the host frontend and scaler input and output
	pads are undefined.</para></listitem>
	<listitem><para>The application configures the frontend input pad resolution to
	2048x1536. The driver propagates the format to the frontend output pad.
	Note that the propagated output format can be different, as in this case,
	than the input format, as the hardware might need to crop pixels (for
	instance when converting a Bayer filter pattern to RGB or YUV).</para></listitem>
	<listitem><para>The application configures the scaler input pad resolution to
	2046x1534 to match the frontend output resolution. The driver propagates
	the format to the scaler output pad.</para></listitem>
	<listitem><para>The application configures the scaler output pad resolution to
	1280x960.</para></listitem>
      </orderedlist>
      </para>

      <para>When satisfied with the try results, applications can set the active
      formats by setting the <structfield>which</structfield> argument to
      <constant>V4L2_SUBDEV_FORMAT_ACTIVE</constant>. Active formats are changed
      exactly as try formats by drivers. To avoid modifying the hardware state
      during format negotiation, applications should negotiate try formats first
      and then modify the active settings using the try formats returned during
      the last negotiation iteration. This guarantees that the active format
      will be applied as-is by the driver without being modified.
      </para>
    </section>

    <section id="v4l2-subdev-selections">
      <title>Selections: cropping, scaling and composition</title>

      <para>Many sub-devices support cropping frames on their input or output
      pads (or possible even on both). Cropping is used to select the area of
      interest in an image, typically on an image sensor or a video decoder. It can
      also be used as part of digital zoom implementations to select the area of
      the image that will be scaled up.</para>

      <para>Crop settings are defined by a crop rectangle and represented in a
      &v4l2-rect; by the coordinates of the top left corner and the rectangle
      size. Both the coordinates and sizes are expressed in pixels.</para>

      <para>As for pad formats, drivers store try and active
      rectangles for the selection targets <xref
      linkend="v4l2-selections-common" />.</para>

      <para>On sink pads, cropping is applied relative to the
      current pad format. The pad format represents the image size as
      received by the sub-device from the previous block in the
      pipeline, and the crop rectangle represents the sub-image that
      will be transmitted further inside the sub-device for
      processing.</para>

      <para>The scaling operation changes the size of the image by
      scaling it to new dimensions. The scaling ratio isn't specified
      explicitly, but is implied from the original and scaled image
      sizes. Both sizes are represented by &v4l2-rect;.</para>

      <para>Scaling support is optional. When supported by a subdev,
      the crop rectangle on the subdev's sink pad is scaled to the
      size configured using the &VIDIOC-SUBDEV-S-SELECTION; IOCTL
      using <constant>V4L2_SEL_TGT_COMPOSE</constant>
      selection target on the same pad. If the subdev supports scaling
      but not composing, the top and left values are not used and must
      always be set to zero.</para>

      <para>On source pads, cropping is similar to sink pads, with the
      exception that the source size from which the cropping is
      performed, is the COMPOSE rectangle on the sink pad. In both
      sink and source pads, the crop rectangle must be entirely
      contained inside the source image size for the crop
      operation.</para>

      <para>The drivers should always use the closest possible
      rectangle the user requests on all selection targets, unless
      specifically told otherwise.
      <constant>V4L2_SEL_FLAG_GE</constant> and
      <constant>V4L2_SEL_FLAG_LE</constant> flags may be
      used to round the image size either up or down. <xref
      linkend="v4l2-selection-flags" /></para>
    </section>

    <section>
      <title>Types of selection targets</title>

      <section>
	<title>Actual targets</title>

	<para>Actual targets (without a postfix) reflect the actual
	hardware configuration at any point of time. There is a BOUNDS
	target corresponding to every actual target.</para>
      </section>

      <section>
	<title>BOUNDS targets</title>

	<para>BOUNDS targets is the smallest rectangle that contains all
	valid actual rectangles. It may not be possible to set the actual
	rectangle as large as the BOUNDS rectangle, however. This may be
	because e.g. a sensor's pixel array is not rectangular but
	cross-shaped or round. The maximum size may also be smaller than the
	BOUNDS rectangle.</para>
      </section>

    </section>

    <section>
      <title>Order of configuration and format propagation</title>

      <para>Inside subdevs, the order of image processing steps will
      always be from the sink pad towards the source pad. This is also
      reflected in the order in which the configuration must be
      performed by the user: the changes made will be propagated to
      any subsequent stages. If this behaviour is not desired, the
      user must set
      <constant>V4L2_SEL_FLAG_KEEP_CONFIG</constant> flag. This
      flag causes no propagation of the changes are allowed in any
      circumstances. This may also cause the accessed rectangle to be
      adjusted by the driver, depending on the properties of the
      underlying hardware.</para>

      <para>The coordinates to a step always refer to the actual size
      of the previous step. The exception to this rule is the source
      compose rectangle, which refers to the sink compose bounds
      rectangle --- if it is supported by the hardware.</para>

      <orderedlist>
	<listitem><para>Sink pad format. The user configures the sink pad
	format. This format defines the parameters of the image the
	entity receives through the pad for further processing.</para></listitem>

	<listitem><para>Sink pad actual crop selection. The sink pad crop
	defines the crop performed to the sink pad format.</para></listitem>

	<listitem><para>Sink pad actual compose selection. The size of the
	sink pad compose rectangle defines the scaling ratio compared
	to the size of the sink pad crop rectangle. The location of
	the compose rectangle specifies the location of the actual
	sink compose rectangle in the sink compose bounds
	rectangle.</para></listitem>

	<listitem><para>Source pad actual crop selection. Crop on the source
	pad defines crop performed to the image in the sink compose
	bounds rectangle.</para></listitem>

	<listitem><para>Source pad format. The source pad format defines the
	output pixel format of the subdev, as well as the other
	parameters with the exception of the image width and height.
	Width and height are defined by the size of the source pad
	actual crop selection.</para></listitem>
      </orderedlist>

      <para>Accessing any of the above rectangles not supported by the
      subdev will return <constant>EINVAL</constant>. Any rectangle
      referring to a previous unsupported rectangle coordinates will
      instead refer to the previous supported rectangle. For example,
      if sink crop is not supported, the compose selection will refer
      to the sink pad format dimensions instead.</para>

      <figure id="subdev-image-processing-crop">
	<title>Image processing in subdevs: simple crop example</title>
	<mediaobject>
	  <imageobject>
	    <imagedata fileref="subdev-image-processing-crop.svg"
	    format="SVG" scale="200" />
	  </imageobject>
	</mediaobject>
      </figure>

      <para>In the above example, the subdev supports cropping on its
      sink pad. To configure it, the user sets the media bus format on
      the subdev's sink pad. Now the actual crop rectangle can be set
      on the sink pad --- the location and size of this rectangle
      reflect the location and size of a rectangle to be cropped from
      the sink format. The size of the sink crop rectangle will also
      be the size of the format of the subdev's source pad.</para>

      <figure id="subdev-image-processing-scaling-multi-source">
	<title>Image processing in subdevs: scaling with multiple sources</title>
	<mediaobject>
	  <imageobject>
	    <imagedata fileref="subdev-image-processing-scaling-multi-source.svg"
	    format="SVG" scale="200" />
	  </imageobject>
	</mediaobject>
      </figure>

      <para>In this example, the subdev is capable of first cropping,
      then scaling and finally cropping for two source pads
      individually from the resulting scaled image. The location of
      the scaled image in the cropped image is ignored in sink compose
      target. Both of the locations of the source crop rectangles
      refer to the sink scaling rectangle, independently cropping an
      area at location specified by the source crop rectangle from
      it.</para>

      <figure id="subdev-image-processing-full">
	<title>Image processing in subdevs: scaling and composition
	with multiple sinks and sources</title>
	<mediaobject>
	  <imageobject>
	    <imagedata fileref="subdev-image-processing-full.svg"
	    format="SVG" scale="200" />
	  </imageobject>
	</mediaobject>
      </figure>

      <para>The subdev driver supports two sink pads and two source
      pads. The images from both of the sink pads are individually
      cropped, then scaled and further composed on the composition
      bounds rectangle. From that, two independent streams are cropped
      and sent out of the subdev from the source pads.</para>

    </section>

  </section>

  &sub-subdev-formats;