Message ID | 20220303175012.358613-1-u.kleine-koenig@pengutronix.de |
---|---|
State | Changes Requested |
Headers | show |
Series | pwm: tegra: Optimize period calculation | expand |
On Thu, Mar 03, 2022 at 06:50:12PM +0100, Uwe Kleine-König wrote: > Dividing by the result of a division looses precision because the result is > rounded twice. E.g. with clk_rate = 48000000 and period = 32760033 the > following numbers result: > > rate = pc->clk_rate >> PWM_DUTY_WIDTH = 187500 > hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns) = 3052 > rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz) = 6144 > > The exact result would be 6142.5061875 and (apart from rounding) this is > found by using a single division. As a side effect is also a tad > cheaper to calculate. > > Also using clk_rate >> PWM_DUTY_WIDTH looses precision. Consider for > example clk_rate = 47999999 and period = 106667: > > mul_u64_u64_div_u64(pc->clk_rate >> PWM_DUTY_WIDTH, period_ns, > NSEC_PER_SEC) = 19 > > mul_u64_u64_div_u64(pc->clk_rate, period_ns, > NSEC_PER_SEC << PWM_DUTY_WIDTH) = 20 > > (The exact result is 20.000062083332033.) > > With this optimizations also switch from round-closest to round-down. Given > that the calculations were non-optimal for quite some time now which > nobody reported as a problem, this is the opportunity to align the driver's > behavior to the requirements of new drivers. (Note however that the > duty_cycle calculation isn't aligned yet.) > > Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de> > --- > drivers/pwm/pwm-tegra.c | 10 +++++----- > 1 file changed, 5 insertions(+), 5 deletions(-) > > diff --git a/drivers/pwm/pwm-tegra.c b/drivers/pwm/pwm-tegra.c > index e5a9ffef4a71..7fc03a9ec154 100644 > --- a/drivers/pwm/pwm-tegra.c > +++ b/drivers/pwm/pwm-tegra.c > @@ -99,7 +99,7 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > int duty_ns, int period_ns) > { > struct tegra_pwm_chip *pc = to_tegra_pwm_chip(chip); > - unsigned long long c = duty_ns, hz; > + unsigned long long c = duty_ns; > unsigned long rate, required_clk_rate; > u32 val = 0; > int err; > @@ -156,11 +156,9 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > pc->clk_rate = clk_get_rate(pc->clk); > } > > - rate = pc->clk_rate >> PWM_DUTY_WIDTH; > - > /* Consider precision in PWM_SCALE_WIDTH rate calculation */ > - hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns); > - rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz); > + rate = mul_u64_u64_div_u64(pc->clk_rate, period_ns, > + (u64)NSEC_PER_SEC << PWM_DUTY_WIDTH); > > /* > * Since the actual PWM divider is the register's frequency divider > @@ -169,6 +167,8 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > */ > if (rate > 0) > rate--; > + else > + return -EINVAL; Can you elaborate on why this is needed? Previously rate == 0 was a valid case and this could still happen with the above calculations. Thierry
On Fri, Apr 22, 2022 at 06:19:03PM +0200, Thierry Reding wrote: > On Thu, Mar 03, 2022 at 06:50:12PM +0100, Uwe Kleine-König wrote: > > Dividing by the result of a division looses precision because the result is > > rounded twice. E.g. with clk_rate = 48000000 and period = 32760033 the > > following numbers result: > > > > rate = pc->clk_rate >> PWM_DUTY_WIDTH = 187500 > > hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns) = 3052 > > rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz) = 6144 > > > > The exact result would be 6142.5061875 and (apart from rounding) this is > > found by using a single division. As a side effect is also a tad > > cheaper to calculate. > > > > Also using clk_rate >> PWM_DUTY_WIDTH looses precision. Consider for > > example clk_rate = 47999999 and period = 106667: > > > > mul_u64_u64_div_u64(pc->clk_rate >> PWM_DUTY_WIDTH, period_ns, > > NSEC_PER_SEC) = 19 > > > > mul_u64_u64_div_u64(pc->clk_rate, period_ns, > > NSEC_PER_SEC << PWM_DUTY_WIDTH) = 20 > > > > (The exact result is 20.000062083332033.) > > > > With this optimizations also switch from round-closest to round-down. Given > > that the calculations were non-optimal for quite some time now which > > nobody reported as a problem, this is the opportunity to align the driver's > > behavior to the requirements of new drivers. (Note however that the > > duty_cycle calculation isn't aligned yet.) > > > > Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de> > > --- > > drivers/pwm/pwm-tegra.c | 10 +++++----- > > 1 file changed, 5 insertions(+), 5 deletions(-) > > > > diff --git a/drivers/pwm/pwm-tegra.c b/drivers/pwm/pwm-tegra.c > > index e5a9ffef4a71..7fc03a9ec154 100644 > > --- a/drivers/pwm/pwm-tegra.c > > +++ b/drivers/pwm/pwm-tegra.c > > @@ -99,7 +99,7 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > > int duty_ns, int period_ns) > > { > > struct tegra_pwm_chip *pc = to_tegra_pwm_chip(chip); > > - unsigned long long c = duty_ns, hz; > > + unsigned long long c = duty_ns; > > unsigned long rate, required_clk_rate; > > u32 val = 0; > > int err; > > @@ -156,11 +156,9 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > > pc->clk_rate = clk_get_rate(pc->clk); > > } > > > > - rate = pc->clk_rate >> PWM_DUTY_WIDTH; > > - > > /* Consider precision in PWM_SCALE_WIDTH rate calculation */ > > - hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns); > > - rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz); > > + rate = mul_u64_u64_div_u64(pc->clk_rate, period_ns, > > + (u64)NSEC_PER_SEC << PWM_DUTY_WIDTH); > > > > /* > > * Since the actual PWM divider is the register's frequency divider > > @@ -169,6 +167,8 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, > > */ > > if (rate > 0) > > rate--; > > + else > > + return -EINVAL; > > Can you elaborate on why this is needed? Previously rate == 0 was a > valid case and this could still happen with the above calculations. If the calculations before the if block result in rate = 0 this means that the requested period is smaller than the minimal possible period. So refusing this setting is part of the switch from something-like-round-closest to round-down. I will send a v2 explaining that in the commit log. Best regards Uwe
diff --git a/drivers/pwm/pwm-tegra.c b/drivers/pwm/pwm-tegra.c index e5a9ffef4a71..7fc03a9ec154 100644 --- a/drivers/pwm/pwm-tegra.c +++ b/drivers/pwm/pwm-tegra.c @@ -99,7 +99,7 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, int duty_ns, int period_ns) { struct tegra_pwm_chip *pc = to_tegra_pwm_chip(chip); - unsigned long long c = duty_ns, hz; + unsigned long long c = duty_ns; unsigned long rate, required_clk_rate; u32 val = 0; int err; @@ -156,11 +156,9 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, pc->clk_rate = clk_get_rate(pc->clk); } - rate = pc->clk_rate >> PWM_DUTY_WIDTH; - /* Consider precision in PWM_SCALE_WIDTH rate calculation */ - hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns); - rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz); + rate = mul_u64_u64_div_u64(pc->clk_rate, period_ns, + (u64)NSEC_PER_SEC << PWM_DUTY_WIDTH); /* * Since the actual PWM divider is the register's frequency divider @@ -169,6 +167,8 @@ static int tegra_pwm_config(struct pwm_chip *chip, struct pwm_device *pwm, */ if (rate > 0) rate--; + else + return -EINVAL; /* * Make sure that the rate will fit in the register's frequency
Dividing by the result of a division looses precision because the result is rounded twice. E.g. with clk_rate = 48000000 and period = 32760033 the following numbers result: rate = pc->clk_rate >> PWM_DUTY_WIDTH = 187500 hz = DIV_ROUND_CLOSEST_ULL(100ULL * NSEC_PER_SEC, period_ns) = 3052 rate = DIV_ROUND_CLOSEST_ULL(100ULL * rate, hz) = 6144 The exact result would be 6142.5061875 and (apart from rounding) this is found by using a single division. As a side effect is also a tad cheaper to calculate. Also using clk_rate >> PWM_DUTY_WIDTH looses precision. Consider for example clk_rate = 47999999 and period = 106667: mul_u64_u64_div_u64(pc->clk_rate >> PWM_DUTY_WIDTH, period_ns, NSEC_PER_SEC) = 19 mul_u64_u64_div_u64(pc->clk_rate, period_ns, NSEC_PER_SEC << PWM_DUTY_WIDTH) = 20 (The exact result is 20.000062083332033.) With this optimizations also switch from round-closest to round-down. Given that the calculations were non-optimal for quite some time now which nobody reported as a problem, this is the opportunity to align the driver's behavior to the requirements of new drivers. (Note however that the duty_cycle calculation isn't aligned yet.) Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de> --- drivers/pwm/pwm-tegra.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) base-commit: ed14d36498c8d15be098df4af9ca324f96e9de74