Re: [RFC PATCH 0/2] IIO: Filtering - how to handle.

[Jonathan Cameron <jic23@cam.ac.uk>]

On 10/06/11 15:01, JohnLM wrote:
> On Tue, 04 Oct 2011 15:23:42 +0100
> Jonathan Cameron <jic23@cam.ac.uk> wrote:
> 
>> On 10/04/11 13:39, JohnLM wrote:
>>> On Fri, 30 Sep 2011 11:17:57 +0100
>>> Jonathan Cameron <jic23@cam.ac.uk> wrote:
>>>
>>>> Hi All,
>>>>
>>>> One big area we have pretty much glossed over so far is devices
>>>> with controllable hardware filters.  This RFC proposes one option
>>>> for how to handle this. For low pass filters at least, the 3db
>>>> point seems the obvious choice as it allows us to gloss over
>>>> exactly what type of filter it is whilst still capturing it's basic
>>>> property of what it lets through.
>>>>
>>>> What do people think?
>>> I have no exact filters in mind but in general since they affect the
>>> readings I think some kind of framework is needed.
>>> Some generic types should be defined, but even when IIO can't make
>>> out what kind of filter it is, user-space app might so it should be
>>> exposed.
>> True.  But how generic do we want and how do we specify it?
>>
>> The version here assumes that knowing it is low pass is enough, but
>> is that true.  Do we need to know for example if it is a Butterworth
>> filter?
> IMO it makes most sense we categorize them by
> basic functionality. Thus making it 'low pass filter' should be good
> enough, but we can (and should) expose it's type.
Hmm. Going to be interesting to document these.  Your Butterworth
etc classes are well known, but there are some nasty hacky things
out there as well.  Lets keep it as a single sysfs type file when we
do it (not suggesting doing this now unless we have a user).
So the butterworth-2, butterworth-4, butterworth-8 etc here.
They are least almost a single conceptual value...
>>
>> If we do, does it need to be in the naming?   If not, should we
>> perhaps have in_voltage_filter_low_pass_type with named filter
>> types?  Perhaps this also has say butterworth-N for N tap butterworth
>> for example? Or should be have filter_low_pass_type and
>> filter_low_pass_taps? Number of taps is often tied up with the 3db
>> point, which would make things a little interesting.  We might have
>> to buffer the 3db request and do a hardware update on any of type,
>> taps, 3db_frequency and sampling frequency changing.  We are already
>> doing that on frequency in the example given here, so not so bad.
>>
>>>
>>> For filters that can be enabled/disabled userspace could even
>>> abstract the filtering routine to use hardware filter when desired
>>> or use software filter when hardware filter is not present or is
>>> unreliable.
>> I'm unconvinced that we want to do software filtering in kernel land.
>> Can't immediately see the point, unless possibly it was used to do
>> data reduction into a buffer.  Worth keeping in mind, but no something
>> I can see happening any time soon.
> I ran a bit ahead of me. I didn't intend to do this kind of processing
> in kernel space. Surely software filters should stay in userspace, I
> don't see the point of this being in kernel space either.
> What I meant was that enough information about filters must be exposed
> so this abstraction could be done in userspace library or/and app.
Agreed.  Though I an inclined to say - add it when needed rather
than anticipate the need. For now the 3db point works for anything I know
of.

Jonathan