Hi Sakari,
On 2017-04-26 00:21, Sakari Ailus wrote:
Hi Marek,
On Thu, Apr 20, 2017 at 01:23:09PM +0200, Marek Szyprowski wrote:
Hi Laurent,
On 2017-04-20 12:25, Laurent Pinchart wrote:
Hi Marek,
(CC'ing Sakari Ailus)
Thank you for the patches.
On Thursday 20 Apr 2017 11:13:36 Marek Szyprowski wrote:
Dear all,
This is an updated proposal for extending EXYNOS DRM API with generic
support for hardware modules, which can be used for processing image data
>from the one memory buffer to another. Typical memory-to-memory operations
are: rotation, scaling, colour space conversion or mix of them. This is a
follow-up of my previous proposal "[RFC 0/2] New feature: Framebuffer
processors", which has been rejected as "not really needed in the DRM
core":
http://www.mail-archive.com/dri-devel@xxxxxxxxxxxxxxxxxxxxx/msg146286.html
In this proposal I moved all the code to Exynos DRM driver, so now this
will be specific only to Exynos DRM. I've also changed the name from
framebuffer processor (fbproc) to picture processor (pp) to avoid confusion
with fbdev API.
Here is a bit more information what picture processors are:
Embedded SoCs are known to have a number of hardware blocks, which perform
such operations. They can be used in paralel to the main GPU module to
offload CPU from processing grapics or video data. One of example use of
such modules is implementing video overlay, which usually requires color
space conversion from NV12 (or similar) to RGB32 color space and scaling to
target window size.
The proposed API is heavily inspired by atomic KMS approach - it is also
based on DRM objects and their properties. A new DRM object is introduced:
picture processor (called pp for convenience). Such objects have a set of
standard DRM properties, which describes the operation to be performed by
respective hardware module. In typical case those properties are a source
fb id and rectangle (x, y, width, height) and destination fb id and
rectangle. Optionally a rotation property can be also specified if
supported by the given hardware. To perform an operation on image data,
userspace provides a set of properties and their values for given fbproc
object in a similar way as object and properties are provided for
performing atomic page flip / mode setting.
The proposed API consists of the 3 new ioctls:
- DRM_IOCTL_EXYNOS_PP_GET_RESOURCES: to enumerate all available picture
processors,
- DRM_IOCTL_EXYNOS_PP_GET: to query capabilities of given picture
processor,
- DRM_IOCTL_EXYNOS_PP_COMMIT: to perform operation described by given
property set.
The proposed API is extensible. Drivers can attach their own, custom
properties to add support for more advanced picture processing (for example
blending).
This proposal aims to replace Exynos DRM IPP (Image Post Processing)
subsystem. IPP API is over-engineered in general, but not really extensible
on the other side. It is also buggy, with significant design flaws - the
biggest issue is the fact that the API covers memory-2-memory picture
operations together with CRTC writeback and duplicating features, which
belongs to video plane. Comparing with IPP subsystem, the PP framework is
smaller (1807 vs 778 lines) and allows driver simplification (Exynos
rotator driver smaller by over 200 lines).
This seems to be the kind of hardware that is typically supported by V4L2.
Stupid question, why DRM ?
Let me elaborate a bit on the reasons for implementing it in Exynos DRM:
1. we want to replace existing Exynos IPP subsystem:
- it is used only in some internal/vendor trees, not in open-source
- we want it to have sane and potentially extensible userspace API
- but we don't want to loose its functionality
2. we want to have simple API for performing single image processing
operation:
- typically it will be used by compositing window manager, this means that
some parameters of the processing might change on each vblank (like
destination rectangle for example). This api allows such change on each
operation without any additional cost. V4L2 requires to reinitialize
queues with new configuration on such change, what means that a bunch of
ioctls has to be called.
What do you mean by re-initialising the queue? Format, buffers or something
else?
In case of compositor use case, the parameter that is being changed most
frequently is source and/or destination rectangle position and/or size.
If you need a larger buffer than what you have already allocated, you'll
need to re-allocate, V4L2 or not.
We also do lack a way to destroy individual buffers in V4L2. It'd be up to
implementing that and some work in videobuf2.
Well if we would use V4l2, buffers will always come as dmabuf objects. There
is a hard limit of the number of buffers that can be imported to v4l2/vb2
queue to get buffer ids. This also limits easy processing of the buffers
in the compositor, because you would need to reinitialize the v4l2 queues
to get new set of v4l2/vb2 buffer ids.
Another thing is that V4L2 is very stream oriented. For most devices that's
fine as a lot of the parameters are not changeable during streaming,
especially if the pipeline is handled by multiple drivers. That said, for
devices that process data from memory to memory performing changes in the
media bus formats and pipeline configuration is not very efficient
currently, largely for the same reason.
The request API that people have been working for a bit different use cases
isn't in mainline yet. It would allow more efficient per-request
configuration than what is currently possible, but it has turned out to be
far from trivial to implement.
- validating processing parameters in V4l2 API is really complicated,
because the parameters (format, src&dest rectangles, rotation) are being
set incrementally, so we have to either allow some impossible,
transitional
configurations or complicate the configuration steps even more (like
calling some ioctls multiple times for both input and output). In the end
all parameters have to be again validated just before performing the
operation.
You have to validate the parameters in any case. In a MC pipeline this takes
place when the stream is started.
Well, in case of v4l2 one would need to stop and restart 'steaming' on
both queues
of mem2mem device just to change some transformation parameters...
3. generic approach (to add it to DRM core) has been rejected:
http://www.mail-archive.com/dri-devel@xxxxxxxxxxxxxxxxxxxxx/msg146286.html
For GPUs I generally understand the reasoning: there's a very limited number
of users of this API --- primarily because it's not an application
interface.
If you have a device that however falls under the scope of V4L2 (at least
API-wise), does this continue to be the case? Will there be only one or two
(or so) users for this API? Is it the case here?
Using a device specific interface definitely has some benefits: there's no
need to think how would you generalise the interface for other similar
devices. There's no need to consider backwards compatibility as it's not a
requirement. The drawback is that the applications that need to support
similar devices will bear the burden of having to support different APIs.
I don't mean to say that you should ram whatever under V4L2 / MC
independently of how unworkable that might be, but there are also clear
advantages in using a standardised interface such as V4L2.
V4L2 has a long history behind it and if it was designed today, I bet it
would look quite different from what it is now.
IMHO V4l2 becomes both a bit over-engineered because of the 'backwards
compatibility' and too limited on the other hand to support really complex
hardware (iirc none of the top mobile Android HW vendors use V4l2 for their
cameras subsystems). This is however a completely separate topic.
4. this api can be considered as extended 'blit' operation, other DRM
drivers
(MGA, R128, VIA) already have ioctls for such operation, so there is also
place in DRM for it
Added LMML to cc.
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
