Message ID  20190220150300.33021enric.balletbo@collabora.com 

State  New 
Headers  show 
Series 

Related  show 
On Wed, 20 Feb 2019 16:03:00 +0100 Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: > From: Gwendal Grignou <gwendal@chromium.org> > > Calculation was copied from IIO_DEGREE_TO_RAD, but offset added to avoid > rounding error is wrong. It should be only half of the divider. > > Fixes: c14dca07a31d ("iio: cros_ec_sensors: add ChromeOS EC Contiguous Sensors driver") > Signedoffby: Gwendal Grignou <gwendal@chromium.org> > Signedoffby: Enric Balletbo i Serra <enric.balletbo@collabora.com> This one is kind of interesting. See below. >  > > drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c  2 + > 1 file changed, 1 insertion(+), 1 deletion() > > diff git a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > index 89cb0066a6e0..600942af9f9c 100644 >  a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > +++ b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > @@ 103,7 +103,7 @@ static int cros_ec_sensors_read(struct iio_dev *indio_dev, > * Do not use IIO_DEGREE_TO_RAD to avoid precision > * loss. Round to the nearest integer. > */ >  *val = div_s64(val64 * 314159 + 9000000ULL, 1000); > + *val = div_s64(val64 * 314159 + 500ULL, 1000); That is only one of two divides going on. Firstly we divide by 1000 here, then we provide it in fractional form which means that the actual value you get from sysfs etc is val/val2. It's this one we are protecting against rounding error on I guess. Now this is even less obviously because it's not 18000 either, but 18000 * 2^CROS_EC_SENSOR_BITS. Which ultimately means neither answer is correct. Hmm. Not totally sure what the right answer actually is.. Jonathan > *val2 = 18000 << (CROS_EC_SENSOR_BITS  1); > ret = IIO_VAL_FRACTIONAL; > break;
Hi Jonathan, On 20/2/19 17:01, Jonathan Cameron wrote: > On Wed, 20 Feb 2019 16:03:00 +0100 > Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: > >> From: Gwendal Grignou <gwendal@chromium.org> >> >> Calculation was copied from IIO_DEGREE_TO_RAD, but offset added to avoid >> rounding error is wrong. It should be only half of the divider. >> >> Fixes: c14dca07a31d ("iio: cros_ec_sensors: add ChromeOS EC Contiguous Sensors driver") >> Signedoffby: Gwendal Grignou <gwendal@chromium.org> >> Signedoffby: Enric Balletbo i Serra <enric.balletbo@collabora.com> > > This one is kind of interesting. See below. > >>  >> >> drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c  2 + >> 1 file changed, 1 insertion(+), 1 deletion() >> >> diff git a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >> index 89cb0066a6e0..600942af9f9c 100644 >>  a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >> +++ b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >> @@ 103,7 +103,7 @@ static int cros_ec_sensors_read(struct iio_dev *indio_dev, >> * Do not use IIO_DEGREE_TO_RAD to avoid precision >> * loss. Round to the nearest integer. >> */ >>  *val = div_s64(val64 * 314159 + 9000000ULL, 1000); >> + *val = div_s64(val64 * 314159 + 500ULL, 1000); > That is only one of two divides going on. Firstly we divide by 1000 here, > then we provide it in fractional form which means that the actual value you get > from sysfs etc is > val/val2. It's this one we are protecting against rounding error on I guess. > Now this is even less obviously because it's not 18000 either, but > 18000 * 2^CROS_EC_SENSOR_BITS. > > Which ultimately means neither answer is correct. Hmm. > Not totally sure what the right answer actually is.. > If I understood well the Gwendal's patch the problem that we're trying to solve is that current calculation is not closer from the float calculation. For 1000dps, the result should be: (1000 * pi ) / 180 >> 15 ~= 0.000532632218 But with current calculation we get $ cat scale 0.000547890 With that patch (modifying the offset to avoid the rounding error) we get a closer result $ cat scale 0.000532631 So, what we're trying to do is have val/val2 closer to the real value. Makes this sense to you or I'm missing something? I can improve the commit message if it's not clear.  Enric > Jonathan > >> *val2 = 18000 << (CROS_EC_SENSOR_BITS  1); >> ret = IIO_VAL_FRACTIONAL; >> break; >
On Fri, 22 Feb 2019 11:24:24 +0100 Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: > Hi Jonathan, > > On 20/2/19 17:01, Jonathan Cameron wrote: > > On Wed, 20 Feb 2019 16:03:00 +0100 > > Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: > > > >> From: Gwendal Grignou <gwendal@chromium.org> > >> > >> Calculation was copied from IIO_DEGREE_TO_RAD, but offset added to avoid > >> rounding error is wrong. It should be only half of the divider. > >> > >> Fixes: c14dca07a31d ("iio: cros_ec_sensors: add ChromeOS EC Contiguous Sensors driver") > >> Signedoffby: Gwendal Grignou <gwendal@chromium.org> > >> Signedoffby: Enric Balletbo i Serra <enric.balletbo@collabora.com> > > > > This one is kind of interesting. See below. > > > >>  > >> > >> drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c  2 + > >> 1 file changed, 1 insertion(+), 1 deletion() > >> > >> diff git a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > >> index 89cb0066a6e0..600942af9f9c 100644 > >>  a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > >> +++ b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c > >> @@ 103,7 +103,7 @@ static int cros_ec_sensors_read(struct iio_dev *indio_dev, > >> * Do not use IIO_DEGREE_TO_RAD to avoid precision > >> * loss. Round to the nearest integer. > >> */ > >>  *val = div_s64(val64 * 314159 + 9000000ULL, 1000); > >> + *val = div_s64(val64 * 314159 + 500ULL, 1000); > > That is only one of two divides going on. Firstly we divide by 1000 here, > > then we provide it in fractional form which means that the actual value you get > > from sysfs etc is > > val/val2. It's this one we are protecting against rounding error on I guess. > > Now this is even less obviously because it's not 18000 either, but > > 18000 * 2^CROS_EC_SENSOR_BITS. > > > > Which ultimately means neither answer is correct. Hmm. > > Not totally sure what the right answer actually is.. > > > > If I understood well the Gwendal's patch the problem that we're trying to solve > is that current calculation is not closer from the float calculation. > > For 1000dps, the result should be: > > (1000 * pi ) / 180 >> 15 ~= 0.000532632218 > > But with current calculation we get > > $ cat scale > 0.000547890 > > With that patch (modifying the offset to avoid the rounding error) we get a > closer result > > $ cat scale > 0.000532631 > > So, what we're trying to do is have val/val2 closer to the real value. Makes > this sense to you or I'm missing something? I can improve the commit message if > it's not clear. I think we are in enough of a mess here with the different dividers that we should just do the maths here, then we can avoid the bia. aiming for nano value. val * pi * 10e12 / (180 * 2^15) div_s64(val * 3141592653000 + 2949120, 5898240) = 532632 vs 532632 for floating point division. Then use IIO_INT_PLUS_NANO to return it. Even then I suspect the +2949120 is only effecting the last digit so you could probably drop it safely enough. I'd certainly rather we had all the magic in one place rather than trying to correct for divisions that aren't apparent here. > >  Enric > > > Jonathan > > > >> *val2 = 18000 << (CROS_EC_SENSOR_BITS  1); > >> ret = IIO_VAL_FRACTIONAL; > >> break; > >
Hi Jonathan, On 3/3/19 17:47, Jonathan Cameron wrote: > On Fri, 22 Feb 2019 11:24:24 +0100 > Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: > >> Hi Jonathan, >> >> On 20/2/19 17:01, Jonathan Cameron wrote: >>> On Wed, 20 Feb 2019 16:03:00 +0100 >>> Enric Balletbo i Serra <enric.balletbo@collabora.com> wrote: >>> >>>> From: Gwendal Grignou <gwendal@chromium.org> >>>> >>>> Calculation was copied from IIO_DEGREE_TO_RAD, but offset added to avoid >>>> rounding error is wrong. It should be only half of the divider. >>>> >>>> Fixes: c14dca07a31d ("iio: cros_ec_sensors: add ChromeOS EC Contiguous Sensors driver") >>>> Signedoffby: Gwendal Grignou <gwendal@chromium.org> >>>> Signedoffby: Enric Balletbo i Serra <enric.balletbo@collabora.com> >>> >>> This one is kind of interesting. See below. >>> >>>>  >>>> >>>> drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c  2 + >>>> 1 file changed, 1 insertion(+), 1 deletion() >>>> >>>> diff git a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >>>> index 89cb0066a6e0..600942af9f9c 100644 >>>>  a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >>>> +++ b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c >>>> @@ 103,7 +103,7 @@ static int cros_ec_sensors_read(struct iio_dev *indio_dev, >>>> * Do not use IIO_DEGREE_TO_RAD to avoid precision >>>> * loss. Round to the nearest integer. >>>> */ >>>>  *val = div_s64(val64 * 314159 + 9000000ULL, 1000); >>>> + *val = div_s64(val64 * 314159 + 500ULL, 1000); >>> That is only one of two divides going on. Firstly we divide by 1000 here, >>> then we provide it in fractional form which means that the actual value you get >>> from sysfs etc is >>> val/val2. It's this one we are protecting against rounding error on I guess. >>> Now this is even less obviously because it's not 18000 either, but >>> 18000 * 2^CROS_EC_SENSOR_BITS. >>> >>> Which ultimately means neither answer is correct. Hmm. >>> Not totally sure what the right answer actually is.. >>> >> >> If I understood well the Gwendal's patch the problem that we're trying to solve >> is that current calculation is not closer from the float calculation. >> >> For 1000dps, the result should be: >> >> (1000 * pi ) / 180 >> 15 ~= 0.000532632218 >> >> But with current calculation we get >> >> $ cat scale >> 0.000547890 >> >> With that patch (modifying the offset to avoid the rounding error) we get a >> closer result >> >> $ cat scale >> 0.000532631 >> >> So, what we're trying to do is have val/val2 closer to the real value. Makes >> this sense to you or I'm missing something? I can improve the commit message if >> it's not clear. > > I think we are in enough of a mess here with the different dividers that we > should just do the maths here, then we can avoid the bia. > > aiming for nano value. > val * pi * 10e12 / (180 * 2^15) > div_s64(val * 3141592653000 + 2949120, 5898240) = 532632 > vs 532632 for floating point division. > Then use IIO_INT_PLUS_NANO to return it. > > Even then I suspect the +2949120 is only effecting the last digit so > you could probably drop it safely enough. > > I'd certainly rather we had all the magic in one place rather than > trying to correct for divisions that aren't apparent here. > Thanks for the clear explanation, yes, this looks better. I'll do some tests and submit a second version. Thanks! >> >>  Enric >> >>> Jonathan >>> >>>> *val2 = 18000 << (CROS_EC_SENSOR_BITS  1); >>>> ret = IIO_VAL_FRACTIONAL; >>>> break; >>> > >
diff git a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c index 89cb0066a6e0..600942af9f9c 100644  a/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c +++ b/drivers/iio/common/cros_ec_sensors/cros_ec_sensors.c @@ 103,7 +103,7 @@ static int cros_ec_sensors_read(struct iio_dev *indio_dev, * Do not use IIO_DEGREE_TO_RAD to avoid precision * loss. Round to the nearest integer. */  *val = div_s64(val64 * 314159 + 9000000ULL, 1000); + *val = div_s64(val64 * 314159 + 500ULL, 1000); *val2 = 18000 << (CROS_EC_SENSOR_BITS  1); ret = IIO_VAL_FRACTIONAL; break;