Message-ID: <20230418-armhole-cartwheel-cee37778a840@wendy>
Date:   Tue, 18 Apr 2023 12:27:33 +0100
From:   Conor Dooley <conor.dooley@...rochip.com>
To:     Uwe Kleine-König 
        <u.kleine-koenig@...gutronix.de>
CC:     Thierry Reding <thierry.reding@...il.com>,
        Daire McNamara <daire.mcnamara@...rochip.com>,
        <linux-kernel@...r.kernel.org>, <linux-pwm@...r.kernel.org>,
        <linux-riscv@...ts.infradead.org>
Subject: Re: [PATCH v16 1/2] pwm: add microchip soft ip corePWM driver

Hey Uwe,

On Tue, Apr 11, 2023 at 06:25:54PM +0200, Uwe Kleine-König wrote:
> On Tue, Apr 11, 2023 at 02:56:15PM +0100, Conor Dooley wrote:
> > On Tue, Apr 11, 2023 at 12:55:47PM +0200, Uwe Kleine-König wrote:
> > > On Tue, Apr 11, 2023 at 09:56:34AM +0100, Conor Dooley wrote:
> > > > Add a driver that supports the Microchip FPGA "soft" PWM IP core.

> > > > +static int mchp_core_pwm_calc_period(const struct pwm_state *state, unsigned long clk_rate,
> > > > +				     u16 *prescale, u16 *period_steps)
> > > > +{
> > > > +	u64 tmp;
> > > > +	u32 remainder;
> > > > +
> > > > +	/*
> > > > +	 * Calculate the period cycles and prescale values.
> > > > +	 * The registers are each 8 bits wide & multiplied to compute the period
> > > > +	 * using the formula:
> > > > +	 *           (prescale + 1) * (period_steps + 1)
> > > > +	 * period = -------------------------------------
> > > > +	 *                      clk_rate
> > > > +	 * so the maximum period that can be generated is 0x10000 times the
> > > > +	 * period of the input clock.
> > > > +	 * However, due to the design of the "hardware", it is not possible to
> > > > +	 * attain a 100% duty cycle if the full range of period_steps is used.
> > > > +	 * Therefore period_steps is restricted to 0xfe and the maximum multiple
> > > > +	 * of the clock period attainable is (0xff + 1) * (0xfe + 1) = 0xff00
> > > > +	 *
> > > > +	 * The prescale and period_steps registers operate similarly to
> > > > +	 * CLK_DIVIDER_ONE_BASED, where the value used by the hardware is that
> > > > +	 * in the register plus one.
> > > > +	 * It's therefore not possible to set a period lower than 1/clk_rate, so
> > > > +	 * if tmp is 0, abort. Without aborting, we will set a period that is
> > > > +	 * greater than that requested and, more importantly, will trigger the
> > > > +	 * neg-/pos-edge issue described in the limitations.
> > > > +	 */
> > > > +	tmp = mul_u64_u64_div_u64(state->period, clk_rate, NSEC_PER_SEC);
> > > > +	if (!tmp)
> > > > +		return -EINVAL;
> > > > +
> > > > +	if (tmp >= MCHPCOREPWM_PERIOD_MAX) {
> > > > +		*prescale = MCHPCOREPWM_PRESCALE_MAX;
> > > > +		*period_steps = MCHPCOREPWM_PERIOD_STEPS_MAX;
> > > > +
> > > > +		return 0;
> > > > +	}
> > > > +
> > > > +	/*
> > > > +	 * There are multiple strategies that could be used to choose the
> > > > +	 * prescale & period_steps values.
> > > > +	 * Here the idea is to pick values so that the selection of duty cycles
> > > > +	 * is as finegrain as possible.
> > > > +	 * This "optimal" value for prescale can be calculated using the maximum
> > > > +	 * permitted value of period_steps, 0xfe.
> > > > +	 *
> > > > +	 *                period * clk_rate
> > > > +	 * prescale = ------------------------- - 1
> > > > +	 *            NSEC_PER_SEC * (0xfe + 1)
> > > > +	 *
> > > > +	 * However, we are purely interested in the integer upper bound of this
> > > > +	 * calculation, so this division should be rounded up before subtracting
> > > > +	 * 1
> > > > +	 *
> > > > +	 *  period * clk_rate
> > > > +	 * ------------------- was precomputed as `tmp`
> > > > +	 *    NSEC_PER_SEC
> > > > +	 */
> > > > +	*prescale = DIV64_U64_ROUND_UP(tmp, MCHPCOREPWM_PERIOD_STEPS_MAX + 1) - 1;
> > > 
> > > If state->period * clk_rate is 765000000001 you get tmp = 765 and then
> > > *prescale = 2. However roundup(765000000001 / (1000000000 * 255)) - 1 is
> > > 3. The problem here is that you're rounding down in the calculation of
> > > tmp. Of course this is constructed because 765000000001 is prime, but
> > > I'm sure you get the point :-)
> > 
> > Hold that thought for a moment..
> 
> OK, so the correction below is to make up for the wrong rounding here.
> I'd like to have that in a comment. Otherwise it wasn't clear to me.

Sure, no problem.

> > > Also we know that tmp is < 0xff00, so we don't need a 64 bit division
> > > here.
> > 
> > Neither here nor below, true.
> > 
> > > > +	/*
> > > > +	 * Because 0xff is not a permitted value some error will seep into the
> > > > +	 * calculation of prescale as prescale grows. Specifically, this error
> > > > +	 * occurs where the remainder of the prescale calculation is less than
> > > > +	 * prescale.
> > > > +	 * For small values of prescale, only a handful of values will need
> > > > +	 * correction, but overall this applies to almost half of the valid
> > > > +	 * values for tmp.
> > > > +	 *
> > > > +	 * To keep the algorithm's decision making consistent, this case is
> > > > +	 * checked for and the simple solution is to, in these cases,
> > > > +	 * decrement prescale and check that the resulting value of period_steps
> > > > +	 * is valid.
> > > > +	 *
> > > > +	 * period_steps can be computed from prescale:
> > > > +	 *                      period * clk_rate
> > > > +	 * period_steps = ----------------------------- - 1
> > > > +	 *                NSEC_PER_SEC * (prescale + 1)
> > > > +	 *
> > > > +	 */
> > > > +	div_u64_rem(tmp, (MCHPCOREPWM_PERIOD_STEPS_MAX + 1), &remainder);
> > > > +	if (remainder < *prescale) {
> > > > +		u16 smaller_prescale = *prescale - 1;
> > > > +
> > > > +		*period_steps = div_u64(tmp, smaller_prescale + 1) - 1;
> > > > +		if (*period_steps < 255) {
> > > > +			*prescale = smaller_prescale;
> > > > +
> > > > +			return 0;
> > > > +		}
> > > > +	}
> > 
> > ...so in your prime case above, we would initially compute a prescale
> > value that is too large, and then wind up hitting the test of the
> > remainder here, thereby realising that the smaller prescale value is a
> > better fit?
> > Perhaps that's not an acceptable way to handle the issue though.
> 
> IMHO it is, but the comment explaining needs some improvement. It should
> also make clear why rounding cannot lead to prescale - 2 being a
> good/better candidate. (I haven't thought it through, maybe needs some
> improvement?) I wonder if the computation can find all improvements
> given that it only used tmp, but not ->period and clk_rate?!

For tmp > 0xff00 it may be off by more than 255, but that isn't possible
at this stage as we've already eliminated the possibility above.

> > > I don't understand that part. It triggers for tmp = 511. So you prefer
> > > 
> > > 	prescale = 1
> > > 	period_steps = 254
> > > 
> > > yielding period = 510 / clkrate over
> > > 
> > > 	prescale = 2
> > > 	period_steps = 170
> > > 
> > > yielding 513 / clkrate. I wonder why.
> 
> I missed the -= 1 in my example. So it's:
> 
> 	It triggers for tmp = 511. So you prefer
> 
> 		prescale = 1
> 		period_steps = 254
> 
> 	yielding period = 510 / clkrate over
> 
> 		prescale = 2
> 		period_steps = 169
> 
> 	yielding 510 / clkrate.
> 
> Here it's obvious that the former is the better one. But I wonder why
> the former isn't found instantly. Wouldn't
> 
> 	*prescale = tmp / (MCHPCOREPWM_PERIOD_STEPS_MAX + 1) - 1
> 
> give a better approximation in general? (Of course with an additional
> check that *prescale >= 0 then.) 
> 
> ... thinking a bit ... yes, I think that's true:
> 
> If you pick *prescale = tmp / (MCHPCOREPWM_PERIOD_STEPS_MAX + 1) - 1,
> then for each period_steps value ≤ 254 we have:
> 
> 	  (*prescale + 1) * (period_steps + 1)
> 	≤ (*prescale + 1) * 255
> 	≤ (tmp // (MCHPCOREPWM_PERIOD_STEPS_MAX + 1)) * 255
> 	≤ (tmp // 255) * 255
> 	≤ (tmp / 255) * 255
> 	= tmp

tmp = 256

*prescale = 256 // (254 + 1) - 1
          ≈ 0

*prescale = 256 // (prescale + 1) - 1
          = 256 / (0 + 1) - 1
	  = 255

That's then gonna give us one of the broken configurations from the
limitations.

tmp = 257

*prescale = 257 // (254 + 1) - 1
          ≈ 0

*prescale = 257 // (prescale + 1) - 1
          = 257 / (0 + 1) - 1
	  = 256
	  = 0 (registers are 8-bit)

And so on...

I'm quite obviously missing something that you may think is obvious
here, but is not immediately clear to me.

> and you can use the maximal period_steps = 0xfe.
> (Only for tmp < 255 that isn't possible, up to you if you refuse these
> or pick a smaller value for period_steps.)
> 
> > Because 513 > 511 & 254 > 170!
> > Is the aim not to produce a period that is less than or equal to that
> > requested? The aim of this driver is to pick a prescale/period_steps
> > combo that satisfies that constraint, while also trying to maximise the
> > "finegrainness" of the duty cycle.
> > The latter should be stated in a comment above.
> 
> ack. I also wonder if such a change breaks the other assumptions a
> consumer might have. I'll spend some more cycles about that in the next
> round :-)
> 
> > > Also tmp = 511 is the only value
> > > where this triggers. There is a mistake somewhere (maybe on my side).
> > 
> > It should trigger for any value 255 * n < x < 256 * n, no?
> > Say for tmp of 767:
> > *prescale = DIV64_U64_ROUND_UP(767, 254 + 1) - 1 = DIV64_U64_ROUND_UP(3.00784...) - 1 = 3
> > remainder = 0.00784.. * (254 + 1) = 2
> > 
> > Am I going nuts? Wouldn't be the first time that I've made a hames of
> > things here, there are 16 versions for a reason after all.
> 
> Ah, I had a bogus break statement in my python test code. So it's me
> who got it wrong.

No worries :)

Pending an explanation of your calculation above, I've gone and done the
rest of these things. I shan't resend until -rc1 or I get an explanation
of your calculation - whichever happens first!

Cheers,
Conor.

Download attachment "signature.asc" of type "application/pgp-signature" (229 bytes)

Powered by blists - more mailing lists