Re: [PATCH] iio: cros_ec: Fix gyro scale calculation

From: Enric Balletbo i Serra
Date: Wed Mar 06 2019 - 06:16:26 EST


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@xxxxxxxxxxxxx> 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@xxxxxxxxxxxxx> wrote:
>>>
>>>> From: Gwendal Grignou <gwendal@xxxxxxxxxxxx>
>>>>
>>>> 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")
>>>> Signed-off-by: Gwendal Grignou <gwendal@xxxxxxxxxxxx>
>>>> Signed-off-by: Enric Balletbo i Serra <enric.balletbo@xxxxxxxxxxxxx>
>>>
>>> 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;
>>>
>
>