[libcamera-devel] [PATCH v4 3/3] libcamera: pipeline: ipu3: Apply the requested test pattern mode

[Hirokazu Honda]

Hi Kieran,

On Tue, Nov 9, 2021 at 7:16 AM Kieran Bingham
<kieran.bingham at ideasonboard.com> wrote:
>
> Quoting Hirokazu Honda (2021-11-04 06:42:52)
> > This enables ipu3 pipeline handler to apply the test pattern mode
> > per frame.
> >
> > Signed-off-by: Hirokazu Honda <hiroh at chromium.org>
> > Reviewed-by: Jean-Michel Hautbois <jeanmichel.hautbois at ideasonboard.com>
> > ---
> >  src/libcamera/pipeline/ipu3/ipu3.cpp | 28 ++++++++++++++++++++++++++--
> >  1 file changed, 26 insertions(+), 2 deletions(-)
> >
> > diff --git a/src/libcamera/pipeline/ipu3/ipu3.cpp b/src/libcamera/pipeline/ipu3/ipu3.cpp
> > index 63cb7f11..cb141ba9 100644
> > --- a/src/libcamera/pipeline/ipu3/ipu3.cpp
> > +++ b/src/libcamera/pipeline/ipu3/ipu3.cpp
> > @@ -59,6 +59,7 @@ public:
> >         void statBufferReady(FrameBuffer *buffer);
> >         void queuePendingRequests();
> >         void cancelPendingRequests();
> > +       void frameStart(uint32_t sequence);
> >
> >         CIO2Device cio2_;
> >         ImgUDevice *imgu_;
> > @@ -78,6 +79,7 @@ public:
> >         std::unique_ptr<ipa::ipu3::IPAProxyIPU3> ipa_;
> >
> >         std::queue<Request *> pendingRequests_;
> > +       std::queue<Request *> processingRequests_;
>
> I'm a bit weary of having requests on multiple queues like this.
> It might be better if we could add some documentation describing the
> purpose of each queue, and what lifetime a request has on each one.
>
>
> Perhaps a banner comment before processingRequests_ to describe briefly
> that the requests will sit on this queue to synchronise setting controls
> at frameStart time? I'm not sure what level will help - but it seems a
> bit vague for 'processingRequests' and 'pendingRequests' otherwise.
>
>
>
> >
> >         ControlInfoMap ipaControls_;
> >
> > @@ -569,6 +571,8 @@ int PipelineHandlerIPU3::configure(Camera *camera, CameraConfiguration *c)
> >         cio2->sensor()->sensorInfo(&sensorInfo);
> >         data->cropRegion_ = sensorInfo.analogCrop;
> >
> > +       cio2->sensor()->setTestPatternMode(controls::draft::TestPatternModeOff);
> > +
>
> This looks like something the CameraSensor should do at the end of initialisation. (CameraSensor::init())
>
>
> >         /*
> >          * Configure the H/V flip controls based on the combination of
> >          * the sensor and user transform.
> > @@ -797,11 +801,13 @@ void PipelineHandlerIPU3::stop(Camera *camera)
> >         int ret = 0;
> >
> >         data->cancelPendingRequests();
> > +       data->processingRequests_ = {};
>
> We add requests to the processingRequests_ during
> queuePendingRequests(), so maybe we should clear them during
> cancelPendingRequests()?
>
> >         data->ipa_->stop();
> >
> >         ret |= data->imgu_->stop();
> >         ret |= data->cio2_.stop();
> > +
> >         if (ret)
> >                 LOG(IPU3, Warning) << "Failed to stop camera " << camera->id();
> >
> > @@ -852,6 +858,8 @@ void IPU3CameraData::queuePendingRequests()
> >
> >                 info->rawBuffer = rawBuffer;
> >
> > +               processingRequests_.push(request);
> > +
> >                 ipa::ipu3::IPU3Event ev;
> >                 ev.op = ipa::ipu3::EventProcessControls;
> >                 ev.frame = info->id;
> > @@ -1131,8 +1139,8 @@ int PipelineHandlerIPU3::registerCameras()
> >                 data->delayedCtrls_ =
> >                         std::make_unique<DelayedControls>(cio2->sensor()->device(),
> >                                                           params);
> > -               data->cio2_.frameStart().connect(data->delayedCtrls_.get(),
> > -                                                &DelayedControls::applyControls);
> > +               data->cio2_.frameStart().connect(data.get(),
> > +                                                &IPU3CameraData::frameStart);
>
> I like having an explicit/clear frameStart event function here.
>

Sorry, I don't get this comment. Could you elaborate?

> >
> >                 /* Convert the sensor rotation to a transformation */
> >                 int32_t rotation = 0;
> > @@ -1423,6 +1431,22 @@ void IPU3CameraData::statBufferReady(FrameBuffer *buffer)
> >         ipa_->processEvent(ev);
> >  }
> >
> > +void IPU3CameraData::frameStart(uint32_t sequence)
> > +{
> > +       if (!processingRequests_.empty()) {
> > +               /* Assumes applying the test pattern mode affects immediately. */
>
> That comment seems misplaced now.
> Perhaps you could add:
>
>                   /*
>                    * Handle controls which are to be set ready for the
>                    * next frame to start.
>                    */
>
> But I'm not sure of the race/sequencing here. I presume at the point we
> have a frameStart event, any controls we apply will take effect for the
> /next/ frame ? This frame has already started...
>
> Or do we expect to set a control here and have it take effect on this
> frame? (That sounds way too racy to be possible).
>

That's good point.. It is difficult to set control in a proper time as
we're not sure when the proper time is.
We also need to think about the effect to the proceeding frames that
are being processed.

Thanks,
-Hiro
> We're going to need to be able to set other controls at this point,
> perhaps from the IPA too, so I expect we'll see some further rework here
> later.
>
> > +               Request *request = processingRequests_.front();
> > +               processingRequests_.pop();
> > +
> > +               int ret = cio2_.sensor()->applyRequestControls(request);
> > +               if (ret)
> > +                       LOG(IPU3, Error) << "Failed applying controls: " << ret;
> > +       }
> > +
> > +       /* Controls that don't affect immediately are applied in delayedCtrls. */
> > +       delayedCtrls_->applyControls(sequence);
> > +}
> > +
> >  REGISTER_PIPELINE_HANDLER(PipelineHandlerIPU3)
> >
> >  } /* namespace libcamera */
> > --
> > 2.33.1.1089.g2158813163f-goog
> >