Re: [PATCH] media: Documentation: Fix frame interval calculation for raw camera sensors

[Jai Luthra]

Quoting Jacopo Mondi (2026-02-25 13:42:35)
> Sorry, one correction
> 
> On Tue, Feb 24, 2026 at 07:22:24PM +0100, Jacopo Mondi wrote:
> > Hi Jai
> >
> > On Tue, Feb 24, 2026 at 10:55:33AM +0530, Jai Luthra wrote:
> > > Hi Jacopo,
> > >
> > > Thanks for the review.
> > >
> > > Quoting Jacopo Mondi (2026-02-23 14:41:02)
> > > > Hi Jai
> > > >
> > > > On Thu, Feb 19, 2026 at 01:20:50PM +0530, Jai Luthra wrote:
> > > > > The previous frame interval formula used analogue crop dimensions. This
> > > > > breaks down for some sensors when binning.
> > > > >
> > > > > For example in imx219 the minimum FLL (frame length in lines) can be
> > > > > lower than the analogue crop height when binning, which would require a
> > > > > negative VBLANK to represent the actual timing. Similarly, imx283 allows
> > > > > a lower minimum HMAX (line length) when doing 2x2 or 3x3 binning than
> > > > > the analogue crop width of the full resolution mode.
> > > > >
> > > > > The CCS specification also describes under section "8.2.6 Line Length
> > > > > and Frame Length" how the horizontal and vertical readout minimums can
> > > > > be different when binning.
> > > > >
> > > > > Replace the formula with the underlying hardware concepts of LLP (line
> > > > > length in pixels) and FLL (frame length in lines). These terms were
> > > > > chosen to match the CCS specification on raw sensors, as it is a cleaner
> > > > > reference compared to a typical sensor vendor datasheet.
> > > > >
> > > > > Finally, define the blanking controls relative to the active pixel
> > > > > readout (post-binning) rather than the analogue crop size. This matches
> > > > > what most sensor drivers already do, and also what applications like
> > > > > libcamera expect. In "Figure 42" of CCS specification too, we see a
> > > > > similar definition:
> > > > >
> > > > >   frame interval = (output width + HBLANK) *
> > > > >                    (output height + VBLANK) / pixel rate
> > > > >
> > > > > Also add a note in the "Writing camera sensor drivers" guide, to ensure
> > > > > this formula is followed by new sensor drivers.
> > > >
> > > > I agree that using the analogue crop rectangle sizes is not correct,
> > > > however, with this new formulation the LLP and FLL values might be
> > > > smaller than the crop rectangles reported through the selection API,
> > > > which I'm not sure it's great from a consistency point of view ?
> > >
> > > While it takes a while to grok, but most sensors I looked at operate in
> > > this way (of reducing the line length or the frame length) when binning,
> > > including CCS (implicitly). A counterexample would really help here.
> > >
> > > >
> > > > Also the below suggested formulation:
> > > >
> > > >         LLP = active width + V4L2_CID_HBLANK
> > > >         FLL = active height + V4L2_CID_VBLANK
> > > >
> > > > Assumes the combined effect of [binning + subsampling] always directly
> > > > affect the readout time of pixels on the pixel array. The CCS specs
> > > > and driver seems to suggest that's not always the case ?
> > >
> > > Why do you think this model can't work for sensors where the readout time is
> > > unaffected?
> > >
> > > Let's say a sensor's pixel array is 1920x1080 with minimum LLP=2000, FLL=1200.
> > >
> > > so, HBLANK minimum = 80, VBLANK minimum = 120
> > >
> > > It also supports a 2x2 binned mode of 960x540 with the same minimum LLP and FLL
> > > (that is same max framerate, no speed up or change in readout).
> > >
> > > so, HBLANK minimum = 2000 - 960 = 1040, VBLANK minimum = 1200 - 540 = 660
> > >
> > > If you match this with the CCS Figure 42 it still makes sense.
> > >
> >
> > Ah well, sure if LLP and FFL remain constant, the usage of the
> > "active sizes" doesn't matter.
> >
> > This means that the min hblank and min vblank have to be doubled to
> > compensate for the halved the active sizes.
> 
> This is of course not correct. blankings doesn't have to be doubled,
> but just enalrged enough to maintain LLP/FFL constant.
> 
> Sorry for the oversight.
> 

No worries, the general idea got across.

> >
> > Speaking in CCS terms, this means the limits reported in Table 86 are
> > always populated.
> >
> > And what I'm struggling with at the moment, is the assumptions that's
> > always be the case for all sensors. Unfortunately I don't have enough
> > experience across all vendors to tell if that's the case..
> >

Indeed, same here. At least the handful of sensors I have dealt with follow
this, so I will go ahead and send a v2 to at least document the current
"Lingua Franca" for blanking controls even if it's as inconsistent as
English :-)

> >
> > > >
> > > > Do we need to distinguish between binning modes that affect the
> > > > timings and binning modes that do not do that?
> > > >
> > > > As we're going to introduce a control for binning to report the
> > > > binning factor in the image dimension domain, should we introduce
> > > > a control to specify the binning factor in the image timing domain ?
> > > >
> > > >         LLP = (analog_crop_width + HBLANK) / binning_timing_h
> > > >         FLL = (analog_crop_height + VBLANK) / binning_timing_v
> > > >         frame_interval = LLP * FLL / pixel_rate
> > > >
> > > > I'm not 100% sure this is correct however, as the blankings should be
> > > > expressed on a different clock domain that the pixel sampling rate,
> > > > but I guess this is a reasonable approximation ?
> > >
> > > What does these two extra controls really offer us?
> > >
> > > All sensors we have seen thus far map their LLP/FLL (or equivalent HTOT/VTOT)
> > > values with respect to the digital readout, and not the analogue pixel array.
> > >
> > > If we add these two controls, we will have to support two different models for
> > > frame interval calculation in the application layer too. Which I'm fine with if
> > > it has a practical benefit, that is, it makes it easier to deal with some
> > > particular sensor.
> >
> > Not sure why you would need two modes to calculate timings in
> > userspace.
> >

As Sakari pointed out, there are in fact 3 models (2 already present in
upstream)

> > >
> > > And if exposing the LLP/FLL directly offer the same benefit, then that is
> > > cleaner, as it leaves the HBLANK/VBLANK as-is in the new model.
> > >
> >
> > As per above, if the consensus is that having the limits updated when
> > binning to maintaine the LLP/FFL values will work for all sensors,
> > then I'm fine with that.
> >

Ack. I think we can discuss that further when common raw sensor model is
being finalized.

> > > Thanks,
> > > Jai
> > >
> > > [...]

Thanks,
    Jai