Re: [bmp280 v1 4/6] iio: pressure: bmp280: enable filter settings for BMx280

[Jonathan Cameron <jic23@kernel.org>]

On Sun, 10 Aug 2025 21:01:24 +0200
ASSI <Stromeko@nexgo.de> wrote:

> Jonathan Cameron writes:
> > On Sun,  3 Aug 2025 16:08:00 +0200
> > Achim Gratz <Achim.Gratz@Stromeko.DE> wrote:
> >  
> >> These devices were using a hardcoded IIR filter of length 4.  Enable
> >> filter_low_pass_3db_frequency settings to control the filter length
> >> settings of the device (as done already for the BMx380 and BMx580
> >> devices, even though the 3dB corner has an inverse relation to the
> >> filter length).  Remove an offset of 1 from the internal handling of
> >> the available values.  
> >
> > This confuses me.  Are we saying those other devices have a 3db frequency
> > control that is not obeying the ABI? If so that sneaked in past me but
> > please don't continue that (and we should fix that broken use of the ABI).  
> 
> This is correct.  The attribute directly controls the tap length of the
> filter and more taps means lower 3dB corner frequency.  Besides, it
> isn't easy to even figure out what the corner frequency is, since it
> depends both on further settings of the device (provided you want the
> actual corner frequency and not one that's normalized to the sampling
> frequency) and the internal IIR equation, although you could sort of
> reverse engineer it from the step response graph in the datasheet since
> it's unlikely to be something complex.  I've had a brief look, but there
> doesn't appear to be a more appropriate attribute that could be used
> within the IIO framework, like timeconstant / tau maybe?
> 

We stuck to 3DB point for these because all most anything else is annoyingly
filter type dependent with no real consistency of meaning.  At least 3dB is
something you can establish if you have a reasonable amount of info about the
filter (or a plot!)
Obviously they are almost always effected by sampling frequency so that maths
can be rather non trivial.

Vs a completely misleading tap length control, an approx 3dB eyeballed
off a plot is an improvement. 

Adding more filter parameters is likely to end up with a very complex
userspace, so if we have any way to get an approx 3dB point, stick to that.
Been way to long since I last did any digital filter stuff, but poking google
ai thing with "IIR filter y_t = (a -1)/a * y_(t-1) + x_t/a"
came back with 3dB in digital angular frequency as

w_c = arccos ((2a^2 - 2a - 1) / (2a(a - 1)) 
Which might be right.

It's pretty unusual to not find a frequency plot that we could read it
directly off on a datasheet :(

Any filter design folk about who can sanity check the AI output?

Jonathan


> 
> Regards,
> Achim.