Hello Geert,

On Thu, Apr 14, 2022 at 12:10:02PM +0200, Geert Uytterhoeven wrote:
> On Wed, Apr 13, 2022 at 10:51 AM Uwe Kleine-König
> <u.kleine-koenig@pengutronix.de> wrote:
> > The newly computed register values are intended to exactly match the
> > previously computed values. The main improvement is that the prescaler
> > is computed directly instead of with a loop. This uses the fact, that
> > prescalers[i] = 1 << (2 * i).
> >
> > Assuming a moderately smart compiler, the needed number of divisions for
> > the case where the requested period is too big, is reduced from 5 to 2.
> 
> I'm not worried about the divisions, but about the ilog2(), which
> uses fls().  The TPU block also exists on SuperH SoCs (although
> currently no SH Linux code has it enabled), and SH uses the fls()
> implementation from asm-generic.
> 
> > --- a/drivers/pwm/pwm-renesas-tpu.c
> > +++ b/drivers/pwm/pwm-renesas-tpu.c
> > @@ -244,7 +244,6 @@ static void tpu_pwm_free(struct pwm_chip *chip, struct pwm_device *pwm)
> >  static int tpu_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm,
> >                           int duty_ns, int period_ns, bool enabled)
> >  {
> > -       static const unsigned int prescalers[] = { 1, 4, 16, 64 };
> >         struct tpu_pwm_device *tpd = pwm_get_chip_data(pwm);
> >         struct tpu_device *tpu = to_tpu_device(chip);
> >         unsigned int prescaler;
> > @@ -254,26 +253,21 @@ static int tpu_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm,
> >         u32 duty;
> >         int ret;
> >
> > -       /*
> > -        * Pick a prescaler to avoid overflowing the counter.
> > -        * TODO: Pick the highest acceptable prescaler.
> > -        */
> >         clk_rate = clk_get_rate(tpu->clk);
> >
> > -       for (prescaler = 0; prescaler < ARRAY_SIZE(prescalers); ++prescaler) {
> > -               period = clk_rate / prescalers[prescaler]
> > -                      / (NSEC_PER_SEC / period_ns);
> > -               if (period <= 0xffff)
> > -                       break;
> > -       }
> > +       period = clk_rate / (NSEC_PER_SEC / period_ns);
> > +       if (period >= 64 * 0x10000 || period == 0)
> > +               return -EINVAL;
> >
> > -       if (prescaler == ARRAY_SIZE(prescalers) || period == 0) {
> > -               dev_err(&tpu->pdev->dev, "clock rate mismatch\n");
> > -               return -ENOTSUPP;
> > -       }
> > +       if (period < 0x10000)
> > +               prescaler = 0;
> > +       else
> > +               prescaler = ilog2(period / 0x10000) / 2 + 1;
> > +
> > +       period >>= 2 * prescaler;
> 
> Although the above is correct, I find it hard to read.
> Hence I'd keep a loop, like:
> 
>     unsigned int prescaler = 0;
>     ...
>     while (period > 0x10000) {
>             period >>= 2;
>             prescalar++;
>     }
> 
> This would even save 2 lines of code ;-)

The "hard to read" part is subjective, I understand it just fine. (But I
admit I wouldn't be surprised if I'm the exception here as I do much
math.) I suggest to judge this by looking at the generated code. I'm not
an expert here (no sh toolchain here, no sh asm foo), but my expectation
is that the compiler notices that 1 <= period / 0x10000 < 64 and then
the inlined fls code should be simplified such that

	ilog2(period / 0x10000) / 2 + 1

simplifies to something like:

	x = period >> 16
	prescaler = 4
	if (!(x & 0xf0u)) {
		x <<= 4;
		prescaler -= 2;
	}
	if (!(x & 0xc0u)) {
		x <<= 2;
		prescaler -= 1;
	}

which I expect to be more efficient than the loop you suggested.

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |