[libcamera-devel] [RFC PATCH] Documentation: IPU3 IPA Design guide

[Laurent Pinchart]

Hi Kieran,

On Tue, Sep 14, 2021 at 11:13:39AM +0100, Kieran Bingham wrote:
> On 14/09/2021 06:37, Laurent Pinchart wrote:
> > On Sun, Sep 12, 2021 at 10:30:17PM +0100, Kieran Bingham wrote:
> >> The IPU3 IPA implements the basic 3a using the ImgU ISP.
> > 
> > s/3a/3A/
> > 
> > Technically speaking we're only implementing 2 out of 3 at this point
> > :-)
> > 
> >> Provide an overview document to describe it's operations, and provide a
> > 
> > s/it's/its/
> > 
> >> block diagram to help visualise how the components are put together to
> >> assist any new developers exploring the code.
> >>
> >> Signed-off-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> >>
> >> ---
> >> This is really just an RFC: Particularly:
> >>
> >>  - Should this content actually live in Documentation?
> >>    or is it more suited to a page under src/ipa/ipu3/...
> > 
> > Good question. We tend to document things in the code where possible.
> > Would there be any drawback from doing so in this case ?
> 
> I don't think so - I think the question mostly comes from
> 'discoverability' of the documentation.
> 
> 
> Which is also a bigger question that we need to make it easier for
> readers to find the documentation that they desire.
> 
> I've come across a great talk on this recently:
>   - What nobody tells you about documentation - Daniele Procida
>   - PyConAu
>   - https://2017.pycon-au.org/schedule/presentation/15/
>   - https://youtu.be/t4vKPhjcMZg
> 
> And that's been spun into a website to document how to document ;-)
>   - https://diataxis.fr/

Thanks, I'll have a look.

> So, I think this could indeed live under src/ipa/ipu3 but it needs to be
> clear that this should be the first thing to read when looking into the IPA.

Agreed. I could envision moving it to Documentation/ later, as part of a
longer document that uses the IPU3 IPA to teach how to write algorithms,
but at that point we'll start writing a book about algorithm design :-)

> >>  - This is not complete to document all functionality
> >>    but gives an overview of the current implementation
> > 
> > Sounds good to me.
> > 
> >>  - Is there anything anyone would find helpful for me to
> >>    extend/write about on this?
> >>
> >>  - It could likely get merged in with some of the work
> >>    that Jean-Michel is doing, and might duplicate some of
> >>    the parts he is documenting, but this was an aim to
> >>    write a single page overview as a getting started with
> >>    the IPU3 IPA ....
> > 
> > I think it's *very* useful to have an overview.
> > 
> >>  .../guides/ipu3-ipa-design-guide.rst          | 144 ++++++++++++++++++
> >>  1 file changed, 144 insertions(+)
> >>  create mode 100644 Documentation/guides/ipu3-ipa-design-guide.rst
> >>
> >> diff --git a/Documentation/guides/ipu3-ipa-design-guide.rst b/Documentation/guides/ipu3-ipa-design-guide.rst
> >> new file mode 100644
> >> index 000000000000..a1d0f13fbb00
> >> --- /dev/null
> >> +++ b/Documentation/guides/ipu3-ipa-design-guide.rst
> >> @@ -0,0 +1,144 @@
> >> +IPU3 IPA Architecture Design and Overview
> >> +=========================================
> >> +
> >> +The IPU3 IPA is built as a modular and extensible framework with an
> >> +upper layer to manage the interactions with the pipeline handler, and
> >> +the image processing algorithms split to compartmentalise the processing
> >> +required for each accellerator cluster provided by the ImgU ISP.
> > 
> > I've commented on the expression "accelerator cluster" when reviewing
> > Jean-Michel's series. I've then seen that it's mentioned in the IPU3
> > kernel documentation in one place, I suppose that's where it comes from.
> > From what I understand, the expression is used to refer to a processing
> > block (or sub-block) that is fully implemented in hardware, as opposed
> > as being implemented partly or fully in the ImgU firmware. I don't think
> > that's relevant from an IPA point of view, it's an internal
> > implementation detail of the ImgU. I'd thus use "processing block" or a
> > similar term instead.
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/staging/media/ipu3/include/uapi/intel-ipu3.h#n2428
> 
> Describes it as:
> 
> """
> ACC refers to the HW cluster containing all Fixed Functions (FFs). Each
> FF implements a specific algorithm.
> """
> 
> So in fact, yes - the term Accelerator Cluster is a 'group' of all the
> hardware accelerator functions, rather than each individual one.
> 
> Each individual function (fixed function) can indeed be better named as
> a processing block.
> 
> 
> Lets make this more like
> 
> """
> The IPU3 IPA is built as a modular and extensible framework with an
> upper layer to manage the interactions with the pipeline handler, and
> the image processing algorithms split to compartmentalise the processing
> required for each processing block, making use of the fixed-function
> accelerators provided by the ImgU ISP.
> """

Sounds good to me. Let's standardize on the same "processing block"
terminology in the IPU3 IPA documentation series too.

> >> +
> >> +The core IPU3 class is responsible for initialisation and construction
> >> +of the algorithm components, processing controls set by the requests
> >> +from applications, and managing events from the Pipeline handler.
> > 
> > s/Pipeline/pipeline/
> > 
> >> +
> >> +::
> >> +
> >> +        ┌───────────────────────────────────────────┐
> >> +        │      IPU3 Pipeline Handler                │
> >> +        │   ┌────────┐    ┌────────┐    ┌────────┐  │
> >> +        │   │        │    │        │    │        │  │
> >> +        │   │ Sensor ├───►│  CIO2  ├───►│  ImgU  ├──►
> >> +        │   │        │    │        │    │        │  │
> >> +        │   └────────┘    └────────┘    └─▲────┬─┘  │    P: Parameter Buffer
> >> +        │                                 │P   │    │    S: Statistics Buffer
> >> +        │                                 │    │S   │
> >> +        └─┬───┬───┬──────┬────┬────┬────┬─┴────▼─┬──┘    1: init()
> >> +          │   │   │      │ ▲  │ ▲  │ ▲  │ ▲      │       2: configure()
> >> +          │1  │2  │3     │4│  │4│  │4│  │4│      │5      3: mapBuffers(), start()
> >> +          ▼   ▼   ▼      ▼ │  ▼ │  ▼ │  ▼ │      ▼       4: processEvent()
> >> +        ┌──────────────────┴────┴────┴────┴─────────┐    5: stop(), unmapBuffers()
> >> +        │ IPU3 IPA                                  │
> >> +        │                 ┌───────────────────────┐ │
> >> +        │ ┌───────────┐   │ Algorithms            │ │
> >> +        │ │IPAContext │   │          ┌─────────┐  │ │
> >> +        │ │ ┌───────┐ │   │          │ ...     │  │ │
> >> +        │ │ │       │ │   │        ┌─┴───────┐ │  │ │
> >> +        │ │ │  SC   │ │   │        │ Tonemap ├─┘  │ │
> >> +        │ │ │       │ ◄───►      ┌─┴───────┐ │    │ │
> >> +        │ │ ├───────┤ │   │      │ AWB     ├─┘    │ │
> >> +        │ │ │       │ │   │    ┌─┴───────┐ │      │ │
> >> +        │ │ │  FC   │ │   │    │ AGC     ├─┘      │ │
> >> +        │ │ │       │ │   │    │         │        │ │
> >> +        │ │ └───────┘ │   │    └─────────┘        │ │
> >> +        │ └───────────┘   └───────────────────────┘ │
> >> +        └───────────────────────────────────────────┘
> >> +         SC: IPASessionConfiguration
> >> +         FC: IPAFrameContext(s)
> >> +
> >> +The IPA instance is constructed and initialised at the point a Camera is
> >> +created by the IPU3 pipeline handler. The initialisation call provides
> 
> "Camera" is capitalised as a libcamera noun. Shouldn't Pipeline Handler
> also be capitalised as it's a libcamera named thing?
> 
> (From your comment above about s/Pipeline/pipeline/. I wonder if instead
> it should be capitalised there and here too.

In Doxygen capitalization generates links to classes, so we have to be
careful about writing "Camera" for instance. When it comes to pipeline
handlers, "Pipeline Handler" won't generate a link, and I wouldn't write
"PipelineHandler" here anyway.

I'm sure there are conventions for all this in documentation. I usually
capitalize words that have special meanings the first time they're
introduced only, but I don't know if that's a good thing to do.

> >> +details about which Camera Sensor is being used, and the controls that
> > 
> > s/Camera Sensor/camera sensor/ ?
> 
> Now that one is awkward, as it's both a libcamera noun, and a real world
> object.
> 
> In this context, I believe camera sensor is correct, (as you've
> suggested) as it's the external world camera sensor details.
> 
> >> +it has available, along with their defaults and ranges.
> > 
> > s/defaults/default values/
> > 
> >> +
> >> +Buffers
> >> +~~~~~~~
> >> +
> >> +The IPA will have Parameter and Statistics buffers shared with it from
> >> +the IPU3 Pipeline handler. These buffers will be passed to the IPA
> >> +before the ``start()`` operation occurs.
> >> +
> >> +The IPA will map the buffers into CPU accessible memory, associated with
> > 
> > s/CPU accessible/CPU-accessible/
> > 
> >> +a buffer ID, and further events for sending or receiving parameter and
> >> +statistics buffers will reference the ID to avoid expensive memory
> >> +mapping operations, or the passing of file handles during streaming.
> >> +
> >> +After the ``stop()`` operation occurs, these buffers will be unmapped,
> >> +and no further access to the buffers is permitted.
> >> +
> >> +Context
> >> +~~~~~~~
> >> +
> >> +Algorithm calls will always have the ``IPAContext`` available to them.
> >> +This context comprises of two parts:
> >> +
> >> +-  IPA Session Configuration
> >> +-  IPA Frame Context
> >> +
> >> +The session configuration structure ``IPASessionConfiguration``
> >> +represents constant parameters determined from either before streaming
> >> +commenced during ``configure()`` or updated parameters when processing
> >> +controls.
> > 
> > I'm not sure about the last part, I thought the session configuration
> > was meant to be constant after configure() ?
> 
> I would expect updates from controls to update the 'active session
> configuration'.
> 
> Of course some of that session configuration will be constant, and not
> change (I don't think we'll change the resolution for instance) - but I
> expect that some parameters will be set such as whether AE is enabled or
> locked.
> 
> That to me is a session configuration parameter. But it can change at
> runtime - and while it remains the same, every new frame will have that
> configuration applied.
> 
> 
> Whereas the FrameContexts are about the context that is handled for each
> frame, and any context shared between algorithms , and is much more dynamic.

This makes me think that we may want to introduce a third structure to
group state that isn't per-frame. I recall designing
IPASessionConfiguration as something that will not change during the
session. Maybe I recall it wrong, but I like the idea of keeping it
static. Separating configuration that is set at the beginning and
doesn't change from controls that are updated dynamically seems a good
design principle to me.

> >> +The IPA Frame Context provides the storage for algorithms for a single
> >> +frame operation.
> >> +
> >> +The ``IPAFrameContext`` structure may be extended to an array, list, or
> >> +queue to store historical state for each frame, allowing algorithms to
> >> +obtain and reference results of calculations which are deeply pipelined.
> >> +This may only be done if an algorithm needs to know the context that was
> >> +applied at the frame the statistics were produced for, rather than the
> >> +previous or current frame.
> >> +
> >> +Presently there is a single ``IPAFrameContext`` without historical data,
> >> +and the context is maintained and updated through successive processing
> >> +operations.
> >> +
> >> +Operating
> >> +~~~~~~~~~
> >> +
> >> +There are three main parts of interactions with the algorithms for the
> >> +IPU3 IPA to operate when running:
> >> +
> >> +-  configure()
> >> +-  processEvent(``EventFillParams``)
> >> +-  processEvent(``EventStatReady``)
> >> +
> >> +The configuration phase allows the pipeline-handler to inform the IPA of
> >> +the current stream configurations, which is then passed into each
> >> +algorithm to provide an opportunity to identify and track state of the
> >> +hardware, such as image size or ImgU pipeline configurations.
> >> +
> >> +Pre-frame preparation
> >> +~~~~~~~~~~~~~~~~~~~~~
> >> +
> >> +When configured, the IPA is notified by the pipeline handler of the
> >> +Camera ``start()`` event, after which incoming requests will be queued
> >> +for processing, requiring a parameter buffer (``ipu3_uapi_params``) to
> >> +be populated for the ImgU. This is passed directly to each algorithm
> >> +through the ``prepare()`` call allowing the ISP configuration to be
> >> +updated for the needs of each component that the algorithm is
> >> +responsible for.
> >> +
> >> +The algorithm should set the use flag (``ipu3_uapi_flags``) for any
> >> +structure that it modifies, and it should take care to ensure that any
> >> +structure set by a use flag is fully initialised to suitable values.
> >> +
> >> +The parameter buffer is returned to the pipeline handler through the
> >> +``ActionParamFilled`` event, and from there queued to the ImgU along
> >> +with a raw frame captured with the CIO2.
> > 
> > This reminds me that we should turn the operations (actions and events)
> > into separate functions in the IPAIPU3Interface. Out of scope for this
> > patch of course.
> 
> You mean directly defining them in Mojo instead of passing them through
> processEvent with an event flag?

Yes.

> I think that would probably be better indeed, I don't see much added
> value to having processEvent() ... (with the small exception that it
> partially shows which functions are called while streaming, but we could
> also document the async/streaming expectations anyway).

processEvent() dates back from when we tried to have a single generic
API for IPAs.

> >> +
> >> +Post-frame completion
> >> +~~~~~~~~~~~~~~~~~~~~~
> >> +
> >> +When the capture of an image is completed, and successfully processed
> >> +through the ImgU, the generated statistics buffer
> >> +(``ipu3_uapi_stats_3a``) is given to the IPA through the
> >> +``EventStatReady`` event. This provides the IPA with an opportunity to
> >> +examine the results of the ISP and run the calculations required by each
> >> +algorithm on the new data. The algorithms may require context from the
> >> +operations of other algorithms, for example, the AWB might choose to use
> >> +a scene brightness determined by the AGC. It is important that the
> >> +algorithms are ordered to ensure that required results are determined
> >> +before they are needed.
> >> +
> >> +The ordering of the algorithm processing is determined by their
> >> +placement in the ``IPU3::algorithms_`` ordered list.
> > 
> > I expect interesting research to happen in this area :-)
> 
> Yes, well it's too early to state any more than that currently I think.

-- 
Regards,

Laurent Pinchart