lkml.org 
[lkml]   [2020]   [Jul]   [15]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v5 07/13] pwm: add support for sl28cpld PWM controller
Hi Uwe,

Am 2020-07-15 18:36, schrieb Uwe Kleine-König:
> On Tue, Jul 14, 2020 at 11:09:28PM +0200, Michael Walle wrote:
>> > My wishlist (just as it comes to my mind, so no guarantee of
>> > completeness):
>> >
>> > - can do 0% duty cycle for all supported period lengths
>> > - can do 100% duty cycle for all supported period lengths
>> > - supports both polarities
>> > - supports immediate change of configuration and after completion of
>> > the currently running period
>> > - atomic update (i.e. if you go from configuration A to configuration B
>> > the hardware guarantees to only emit periods of type A and then type
>> > B. (Depending on the item above, the last A period might be cut off.)
>>
>> We actually discussed this, because the implementation would be
>> easier. But
>> if the change takes place immediately you might end up with a longer
>> duty
>> cycle. Assume the PWM runs at 80% duty cycle and starts with the
>> on-period.
>> If you now change that to 50% you might end up with one successive
>> duty
>> cycle of "130%". Eg. the 80% of the old and right after that you
>> switch to
>> the new 50% and then you'd have a high output which corresponds to a
>> 130%
>> cycle. I don't know if that is acceptable for all applications.
>
> I thought this is a "change takes place immediately" implementation?!
> So
> these problems are actually real here. (And this not happening is
> exactly
> my wish here. Is there a mis-understanding?)

I wasn't talking about the sl28cpld btw. What is the difference between
your proposed "change take place immediately" and "after the cycle".
I understand how the after the cycle should work. But how would the
immediate change work in your ideal PWM?

>> > > > If you change only cycle but not mode, does the hardware complete the
>> > > > currently running period?
>> > >
>> > > No it does not.
>> >
>> > Please document this as a Limitation.
>>
>> I've discussed this internally, for now its a limitation. In the worst
>> case you'd do one 100% duty cycle. Maybe we can fix the hardware. I
>> acknowledge that this is a severe limitation, esp. if you use the PWM
>> for controlling stuff (for now its only LCD backlight.. so thats ok).
>
> That happens if you reduce the duty cycle from A to B and the counter
> is
> already bigger than B but smaller than A, right?
That is correct.

> The fix would be to
> compare for counter >= match instead of counter = match. (Which then
> would result in a period with a duty cycle bigger than B but smaller
> than A. Also not ideal, but probably better.)

This would actually work. I could imagine that comparing "less than"
will
take up more space again. But it would be worth a try; see also below.

>> > > > What about disable()?
>> > >
>> > > Mhh well, it would do one 100% cycle.. mhh ;) Lets see if there we can
>> > > fix that (in hardware), not much we can do in the driver here. We are
>> > > _very_ constraint in size, therefore all that little edge cases fall
>> > > off
>> > > the table.
>> >
>> > You're saying that on disable the hardware emits a constant high level
>> > for one cycle? I hope not ...
>>
>> Mh, I was mistaken, disabling the PWM will turn it off immediately,
>> but
>
> And does turn off mean, the output gets inactive?
> If so you might also disable the hardware if a 0% duty cycle is
> configured assuming this saves some energy without modifying the
> resulting wave form.

Disabling it has some side effects like switching to another function
for this multi function pin. So I'd rather keep it on ;)

>> one 100% duty cycle may happen if you change from a higher to a lower
>> duty cycle setting. See above.
>>
>> > I never programmed a CPLD to emulate a hardware PWM, but I wonder if
>> > these are really edge cases that increase the size of the binary?!
>>
>> At the moment there is only one 8bit register which stores the value
>> which is used for matching. If you want to change that setting after
>> a whole cycle, you'd use another 8bit register to cache the new value.
>> So this would at least needs 8 additional flip-flops. This doesn't
>> sound much, but we are already near 100% usage of the CPLD. So its
>> hard to convince people why this is really necessary.
>
> OK. (Maybe there is enough space to allow implementing 100% for mode
> 0?)

Little bit here a little bit there ;) TBH there are some more critical
bugs which would need to be fixed first. So this would need to be a
limitation for now.

-michael

\
 
 \ /
  Last update: 2020-07-15 19:46    [W:0.085 / U:0.676 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site