Use the PixelFormat instead of unsigned int where a pixel format is to
be used. PixelFormat is defined as an unsigned int but is about to be
turned into a class to add functionality.
There is no functional change in this patch.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Associate to each Camera a ControlList which contains the Camera
properties as created by pipeline handlers in the pipeline handler's
CameraData and provide an operation to retrieve them.
Collect properties from the camera sensor in all pipeline handlers that
support one (IPU3, RKISP1 and VIMC).
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
rkisp1 kernel driver was merged upstream with minor changes in the
topology from the original driver libcamera based it's first support to
rkisp1.
Adapt libcamera pipeline to work with upstream driver.
* Remove subdevice dphy from the pipeline.
* Add resizer in the pipeline.
* Fix links.
* Update entity names.
Signed-off-by: Helen Koike <helen.koike@collabora.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
When inserting an element with emplace(), the element is constructed
in-place with the parameters to the emplace() method being forwarded to
the constructor of the element. For std::map containers, the element is
an std::pair<const Key, T>. The constructors of std::pair<T1, T2> fall
into three categories:
(1) Default, copy and move constructors (and related versions)
(2) Constructors that take lvalue or rvalue references to T1 and T2
(3) A forwarding constructor that forwards parameters to the
constructors of T1 and T2
The first category isn't useful in most cases for std::map::emplace(),
as the caller usually doesn't have an existing std::pair<const Key, T>
for the element to be inserted.
The constructor from the third category is useful to avoid constructing
intermediate Key or T instances when the caller doesn't have them
available. This constructor takes two std::tuple arguments that contain
the arguments for the Key and T constructors, respectively. Due to
template deduction rules, usage of such a constructor couldn't be
deduced by the compiler automatically in all cases, so the constructor
takes a first argument of type std::piecewise_construct_t that lets the
caller force the usage ot the forwarding constructor (also known for
this reason as the piecewise constructor). The caller uses a construct
such as
map.emplace(std::piecewise_construct,
std::forward_as_tuple(args_for_Key, ...),
std::forward_as_tuple(args_for_T, ...));
This syntax is a bit heavy, but is required to construct Key and T
in-place from arguments to their non-default constructor (it is also the
only std::pair non-default constructor that can be used for non-copyable
non-movable types).
When the caller of std::map::emplace() already has references to a Key
and a T, they can be passed to the std::pair piecewise constructor, and
this will create std::tuple instance to wrap the Key and T references
arguments to ultimately pass them to the Key and T copy constructors.
map.emplace(std::piecewise_construct,
std::forward_as_tuple(Key_value),
std::forward_as_tuple(T_value));
While this mechanism works, it's unnecessary complex. A constructor of
std::pair that takes references to Key and T can be used without any
performance penalty, as it will also call the copy constructor of Key
and T. In this case we can use a simpler constructor of std::pair, and
thus a simpler call of std::map::emplace.
map.emplace(Key_value, T_value);
We have a couple occurrences of this above misuse of piecewise
construction. Simplify them, which simplifies the code and reduces the
generated code size.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Kieran Bingham <kieran.bingham@ideasonboard.com>
With the FrameBuffer interface in place there is no need for the Camera
to call into the specific pipelines allocation and freeing of buffers as
it no longer needs to be synchronized with buffer allocation by the
application.
Remove the function prototypes in the pipeline handler base class and
fold the functionality in the pipelines start() and stop() functions
where needed. A follow up patch will remove the now no-op
Camera::allocateBuffers() and Camera::freeBuffers().
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The Buffer interface is no longer in use and can be removed. While doing
so clean up the two odd names (dequeueFrameBuffer() and
queuedFrameBuffers_) that had to be used when adding the FrameBuffer
interface.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Switch to the FrameBuffer interface where all buffers are treated as
external buffers and are allocated outside the camera. Applications
allocating buffers using libcamera are switched to use the
FrameBufferAllocator helper.
Follow-up changes to this one will finalize the transition to the new
FrameBuffer interface by removing code that is left unused after this
change.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Extend the pipeline handlers to support the FrameBuffer API with three
new methods to handle allocation, importing and freeing of buffers. The
new methods will replace allocateBuffers() and freeBuffers().
The FrameBuffer API will use the methods on a stream level and either
allocate or import buffers for each active stream controlled from the
Camera class and an upcoming FrameBufferAllocator helper. With this new
API the implementation in pipeline handlers can be made simpler as all
streams don't need to be handled in allocateBuffers().
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The V4L2VideoDevice class can now operate using a FrameBuffer interface,
switch the IPU3 CIO2 and statistics buffer to use it. We can not convert
the application-facing buffers yet.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The V4L2VideoDevice class can now operate using a FrameBuffer interface,
switch the RkISP1 statistics and parameters buffer to use it. We can not
convert the application-facing buffers yet.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
It's common for applications to create and queue a new request in a
previous request completion handler. When the new request gets queued to
the RkISP1 pipeline handler it tries to find a parameters and statistic
buffer to be used with the request. The problem is if the pipeline depth
is already filled there are no internal buffers free to be used by the
new request.
This was solved by allocation one more parameters and statistic buffer
then the pipeline depth, this is waste full. Instead free the resources
of the request that has completed before it is signaled to the
application, this way if the pipeline depth is full it can reuse the
internal resources and the wasteful allocation can be removed.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Move the metadata retrieved when dequeuing a V4L2 buffer into a
FrameMetadata object. This is done as a step to migrate to the
FrameBuffer interface as the functions added to Buffer around
FrameMetadata match the ones in FrameBuffer.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
It is not libcamera's responsibility to handle memory mappings. Switch
from the soon to be removed Plane class which deals with memory
mappings to FrameBuffer::Plane which just describes it. This makes the
transition to the full FrameBuffer easier.
As the full FrameBuffer interface has not yet spread to all parts of
libcamera core it is hard to create efficient caching of memory mappings
in the qcam application. This will be fixed in a later patch, for now
the dmabuf is mapped and unmapped each time it is seen by the
application.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Switch the IPA interfaces and implementations to use the Framebuffer
interface.
- The IPA interface is switched to use the simpler FrameBuffer::Plane
container when carrying dmabuf descriptions (fd and length) over the
pipeline/IPA boundary.
- The RkISP1 IPA implementation takes advantage of the new simpler and
safer (better control over file descriptors) FrameBuffer interface.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Expecting pipeline handler implementations of queueRequest() to call
the base class queueRequest() at the correct point have led to different
behaviors between the pipelines.
Fix this by splitting queueRequest() into a base class implementation
which handles the bookkeeping and a new queueRequestDevice() that is
to be implemented by pipeline handlers and only deals with committing the
request to the device.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The IPAInterface::configure() operation receives a map of ControlInfoMap
instances. Pass const references instead to avoid copies when not
required (the callee can still make manual copies), and to allow for the
future serialization layer to keep references to the original object.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
In preparation for serialization, index the ControlList by unsigned int.
This will allow deserializing a ControlList without requiring external
information.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Fill the StreamConfiguration with all supported formats. The list of
supported formats is currently hardcoded based on the limits of the vimc
driver.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
The V4L2ControlList class only provides a convenience constructor for
the ControlList, which can easily be moved to the ControlList class and
may benefit it later (to construct a ControlList from controls supported
by a camera). Move the constructor and remove V4L2ControlList.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
The ControlInfoMap and V4L2ControlInfoMap classes are very similar, with
the latter adding convenience accessors based on numerical IDs for the
former, as well as a cached idmap. Both features can be useful for
ControlInfoMap in the context of serialisation, and merging the two
classes will further simplify the IPA API.
Import all the features of V4L2ControlInfoMap into ControlInfoMap,
turning the latter into a real class. A few new constructors and
assignment operators are added for completeness.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
V4L2ControlRange only offers a convenience constructor for a
ControlRange. Store the ControlRange instead of V4L2ControlRange in
V4L2ControlInfoMap to make the map less dependent on V4L2 types.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
The V4L2ControlInfo class only stores a ControlRange. Make it inherit
from ControlRange to provide a convenience constructor from a struct
v4l2_query_ext_ctrl and rename it to V4L2ControlRange.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
To bring the libcamera and V4L2 control info maps closer, index the
latter by ControlId * like the former. As V4L2ControlInfoMap is widely
indexed by V4L2 numerical IDs, add accessors based on numerical IDs.
This allows complete removal of the ControId pointer from the
V4L2ControlInfo, as the ControId is accessible as the key when iterating
over the map. A handful of users have to be modified to adapt to the
change.
The controlInfo argument from V4L2Device::updateControls() can also be
removed as it itsn't used anymore.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Now that the V4L2ControlList is merely a helper to construct a
ControlList for V4L2 controls, without any data member, all controls can
be transferred between pipeline handlers and IPAs using ControlList
only. Remove the v4l2controls member for IPAOperationData and use the
control member instead.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Tested-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The V4L2Device class uses V4L2ControlList as a controls container for
the getControls() and setControls() operations. Having a distinct
container from ControlList will makes the IPA API more complex, as it
needs to explicitly transport both types of lists. This will become even
more painful when implementing serialisation and deserialisation.
To simplify the IPA API and ease the implementation of serialisation and
deserialisation, replace usage of V4L2ControlList with ControlList in
the V4L2Device (and thus CameraSensor) API. The V4L2ControlList class
becomes a thin wrapper around ControlList that slightly simplifies the
creation of control lists for V4L2 controls, and may be removed in the
future.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Tested-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
With gcc 5 and 6, insertion in a std::multimap copies the pair passed as
an argument to the insert() method. As the mapped type is a non-copyable
std::unique_ptr<>, this fails to compile.
Compilation with newer gcc versions succeed due to support for C++-17
and the fix described in https://cplusplus.github.io/LWG/issue2354. To
support gcc 5 and 6, fix the issue by using std::multimap::emplace().
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Add the plumbing to the pipeline handler to interact with an IPA module.
This change makes the usage of an IPA module mandatory for the rkisp1
pipeline.
The RkISP1 pipeline handler makes use of a timeline component to
schedule actions. This might be useful for other pipeline handlers going
forward so keep the generic timeline implementation separate to make it
easy to break out.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Acked-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
The IPA acts on a camera and not on a pipeline which can expose more
then one camera. Move the IPA reference to the CameraData.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Use the ControlRange class to express the range of a V4L2 control,
replacing the open-coded minimum and maximum fields in the
V4L2ControlInfo class.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Use the ControlValue class to replace the manually crafted data storage
in V4L2Control. This will help sharing code when more data types will be
supported.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The ControlInfo id member is only used in the toString() method of the
class, and nowhere else externally. The same way that ControlValue
doesn't store a ControlId, ControlInfo shouldn't. Remove it.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Rework the control-related classes to improve the API towards
applications. The goal is to enable writing code similar to
Request *req = ...;
ControlList &controls = req->controls();
controls->set(controls::AwbEnable, false);
controls->set(controls::ManualExposure, 1000);
...
int32_t exposure = controls->get(controls::ManualExposure);
with the get and set operations ensuring type safety for the control
values. This is achieved by creating the following classes:
- Control defines controls and is the main way to reference a control.
It is a template class to allow methods using it to refer to the
control type.
- ControlId is the base class of Control. It stores the control ID, name
and type, and can be used in contexts where a control needs to be
referenced regardless of its type (for instance in lists of controls).
This class replaces ControlIdentifier.
- ControlValue is kept as-is.
The ControlList class now exposes two template get() and set() methods
that replace the operator[]. They ensure type safety by infering the
value type from the Control reference that they receive.
The main way to refer to a control is now through the Control class, and
optionally through its base ControlId class. The ControlId enumeration
is removed, replaced by a list of global Control instances. Numerical
control IDs are turned into macros, and are still exposed as they are
required to communicate with IPAs (especially to deserialise control
lists). They should however not be used by applications.
Auto-generation of header and source files is removed for now to keep
the change simple. It will be added back in the future in a more
elaborate form.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Make the control API more explicit when dealing with integer controls by
specifying the size. We already do so for 64-bit integers, using int64_t
and ControlTypeInteger64, do the same for 32-bit integers.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The ControlValue get accessors are implemented with functions of
different names, whlie the set accessors use polymorphism to support
different control types. This isn't very consistent and intuitive. Make
the API clearer by using template methods. This will also have the added
advantage that support for the new types will only require adding
template specialisations, without adding new methods.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
The ipa includes are located in include/libcamera/ipa/. This gives an
incorrect impression that they are a sub-part of the rest of the
libcamera API, while they are the API towards the IPA the same way that
include/libcamera/ contains the API towards applications. To clarify
this, move them to include/ipa/.
The IPA headers are however still part of libcamera, so installing them
to ${prefix}/include/ipa/ would make little sense. To fix this, move the
application facing API to ${prefix}/include/libcamera/libcamera/ when
installed, and the IPA to ${prefix}/include/libcamera/ipa/. When major
versions of libcamera will be released, they could then be installed
side by side in ${prefix}/include/libcamera-${version}/.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Jacopo Mondi <jacopo@jmondi.org>
Linux commit 85ab1aa1fac17bcd ("media: vimc: deb: fix default sink bayer
format") which is part of v5.2 changes the default media bus format for
the debayer subdevices. This leads to a -EPIPE error when trying to use
the raw capture video device nodes.
Fix this by moving the vimc pipeline to use the RGB/YUV Capture capture
video node. As a consequence of this change the scaler in the vimc
pipeline is used and a hard coded upscale of 3 is present in the video
pipeline. This limits the sizes exposed and accepted by libcamera to
multiples of 3.
Update the buffer import test case to match this new format constraint
by setting the resolution to 1920x1080, a resolution which works with
both vimc and vivid. With this new resolution the buffer import test
takes a bit more time to complete 60 buffers and it's runtime also needs
to be increased.
The raw capture video node still needs to be handled by the pipeline as
its format needs to be updated to allow the pipeline format validation
to pass.
Signed-off-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
ISP output pad should be set to YUYV8_2X8 for non-bayer output format.
Bayer formats are not listed in RkISP1CameraConfiguration::validate(),
only non-bayer are listed, so we can set YUYV8_2X8 directly.
This will need to be changed if we add support for bayer output with
libcamera.
Signed-off-by: Helen Koike <helen.koike@collabora.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
When a video device is stopped all the buffers there queued are released
and their state is set to BufferCancelled.
Currently, on buffer completion, cancelled buffers are blindly re-queued
to the ImgU input or CIO2 output devices, preventing them to be
re-started succesfully in future capture sessions.
Fix that by inspecting the buffers status and skip re-queueing if
they're reported as cancelled. For the ImgU output buffer this is not
required, as cancelled request should be reported to applications in
order to report them failure of the capture operations.
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
With the current IPU3 kernel driver implementation, a linked pipe shall be
used (buffers should be queued on it) in order not to block all other pipes.
Currently all links on the ImgU device are only disabled at match() time,
implying that once an ImgU pipe gets linked, it should be used until the
whole pipeline is not re-matched and links disabled again. This is a severe
limitation for applications that wants to switch between cameras using
different pipes going through a full library tear-down and reload.
Perform link disabling at configure() time as well, so that a camera
configuration operation always unlock the usage of the assigned pipe,
regardless of the previously linked ones.
Unfortunately this requires a camera start/stop sequence to always go
through a configure step, a requirement that is not enforced by the
Camera state machine.
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
The internal buffers between the CIO2 and ImgU are freed by the
CIO2Device::stop() method, which is called first when stopping
streaming. The ImgUDevice::stop() method is then called, and attempts to
report completion for all queued buffers, which we have just freed. The
use-after-free corrupts memory, leading to crashes.
Fix this by moving the vector of internal buffers to the IPU3CameraData
where it belongs, and free the buffers after stopping both devices.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
To avoid printing debugging messages related to stream configuration
adjustement when generating a new configuration, set the pixel format
explicitly instead of relying on the internal validate() call to do so.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Add support for importing external buffers in all pipeline handlers.
Use the stream memory type in the pipeline handlers during buffer
allocation to import buffers to or export buffers from the video device.
Signed-off-by: Jacopo Mondi <jacopo@jmondi.org>
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Access the number of allocated buffer for the streams through the stream
configuration instead of the stream's buffers pool.
Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Niklas Söderlund <niklas.soderlund@ragnatech.se>
Niklas Söderlund
Jacopo Mondi
Helen Koike
Laurent Pinchart
Paul Elder