On Fri, Apr 07, 2017 at 03:04:00PM +0530, Laxman Dewangan wrote: > The rate of the PWM calculated as follows: > hz = NSEC_PER_SEC / period_ns; > rate = (rate + (hz / 2)) / hz; > > This has the precision loss in lower PWM rate. > > Change this to have more precision as: > hz = DIV_ROUND_CLOSEST_ULL(NSEC_PER_SEC * 100, period_ns); > rate = DIV_ROUND_CLOSEST(rate * 100, hz) > > Example: > 1. period_ns = 16672000, PWM clock rate is 200KHz. > Based on old formula > hz = NSEC_PER_SEC / period_ns > = 1000000000ul/16672000 > = 59 (59.98) > rate = (200K + 59/2)/59 = 3390 > > Based on new method: > hz = 5998 > rate = DIV_ROUND_CLOSE(200000*100, 5998) = 3334 > > If we measure the PWM signal rate, we will get more accurate period > with rate value of 3334 instead of 3390. > > 2. period_ns = 16803898, PWM clock rate is 200KHz. > Based on old formula: > hz = 59, rate = 3390 > Based on new formula: > hz = 5951, rate = 3360 > > The PWM signal rate of 3360 is more near to requested period than 3333. > > Signed-off-by: Laxman Dewangan <[email protected]> > > --- > Changes from v1: > - None > > Changes from V2: > - Fix the commit message with exact formula used. > --- > drivers/pwm/pwm-tegra.c | 7 +++++-- > 1 file changed, 5 insertions(+), 2 deletions(-) > > diff --git a/drivers/pwm/pwm-tegra.c b/drivers/pwm/pwm-tegra.c > index 0a688da..21518be 100644 > --- a/drivers/pwm/pwm-tegra.c > +++ b/drivers/pwm/pwm-tegra.c > @@ -76,6 +76,7 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct > pwm_device *pwm, > struct tegra_pwm_chip *pc = to_tegra_pwm_chip(chip); > unsigned long long c = duty_ns; > unsigned long rate, hz; > + unsigned long long ns100 = NSEC_PER_SEC; > u32 val = 0; > int err; > > @@ -94,9 +95,11 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct > pwm_device *pwm, > * cycles at the PWM clock rate will take period_ns nanoseconds. > */ > rate = clk_get_rate(pc->clk) >> PWM_DUTY_WIDTH; > - hz = NSEC_PER_SEC / period_ns; > > - rate = (rate + (hz / 2)) / hz; > + /* Consider precision in PWM_SCALE_WIDTH rate calculation */ > + ns100 *= 100; > + hz = DIV_ROUND_CLOSEST_ULL(ns100, period_ns);
I think hz could overflow for small enough values of period_ns. I've sent a patch that makes hz unsigned long long. While at it, the patch also removes the ns100 variable which isn't really necessary here. Thierry
signature.asc
Description: PGP signature
