Re: [PATCH v2 2/2] iio: light: Add support for APDS9306 Light Sensor

[Subhajit Ghosh <subhajit.ghosh@tweaklogic.com>]

On 31/10/23 17:41, Matti Vaittinen wrote:
> On 10/30/23 12:21, Matti Vaittinen wrote:
>> Hi dee Ho peeps,
>>
>> On 10/29/23 17:51, Matti Vaittinen wrote:
>>> On 10/28/23 18:20, Jonathan Cameron wrote:
>>>> On Fri, 27 Oct 2023 18:15:45 +1030
>>>> Subhajit Ghosh <subhajit.ghosh@tweaklogic.com> wrote:
>>>>
>>>>> Driver support for Avago (Broadcom) APDS9306 Ambient Light Sensor with als
>>>>> and clear channels with i2c interface. Hardware interrupt configuration is
>>>>> optional. It is a low power device with 20 bit resolution and has
>>>>> configurable adaptive interrupt mode and interrupt persistence mode.
>>>>> The device also features inbuilt hardware gain, multiple integration time
>>>>> selection options and sampling frequency selection options.
> 
> ...
> 
>>>>> +static int apds9306_scale_set(struct apds9306_data *data, int val, int val2)
>>>>> +{
>>>>> +    int i, ret, time_sel, gain_sel;
>>>>> +
>>>>> +    /* Rounding up the last digit by one, otherwise matching table fails! */
>>>>
>>>> Interesting.  Sounds like a question for Matti?
>>>
>>> Sounds odd indeed. I assume this happens when scale setting is requested using one of the exact values advertised by the available scales from the GTS? This does not feel right and the +1 does not ring a bell to me. I need to investigate what's going on. It would help if you could provide the values used as val and val2 for the setting.
>>>
>>> This will take a while from me though - I'll try to get to this next week. Thanks for pointing out the anomaly!
>>>
>>
>> I think I have a rough understanding. I did a Kunit test which goes through all the available scales values from the gts->avail_all_scales_table and all integration times, and feeds them to the logic below. It seems the first culprit is hit by:
>> val = 0, val2 = 125025502.
>>
>> Problem is that the 125025502 is rounded. The exact linearized NANO scale resulting from time multiplier 128, gain multiplier 1 is 125025502.5 - which means we will see rounding.
>>
>>>>
>>>>> +    if (val2 % 10)
>>>>> +        val2 += 1;
>>
>> For a while I was unsure if this check works for all cases because I see linearized scales:
>> 250051005 - multipliers 1x, 64x
>> 83350335 - multipliers 3x, 64x and 6x, 32x
>> 27783445 - multipliers 9x, 64x.
>>
>> For those we will get + 1 added to val2 even though there is no rounding. It appears this is not a problem because the iio_gts_get_gain() (which is used to figure out the required total gain to get the desired scale) does not require the scale to be formed by exact multiples of gain.
> 
> ...
> 
>> I think it would be very nice if the gts-helpers could do the rounding when computing the available scales, but that'd require some thinking. Fixup patch is still very welcome ;)
> 
> I did some further experimenting. Basically, I did a "hack" which always rounds up the available-scales values if division results a remainder. This way the values advertised by the available_scales did find the matching table.
> 
> It is a tiny bit icky because for example the scale 6945861.25 becomes 6945862 in available-scales. Also, I assume that if we "hack" just the available-scales and don't fix the rest of the logic, setting 6945862 will read back as 6945861 (I haven't tested this though). Also, the 20837583.75 will be 20837583 in available-scales but 20837582 when read back, resulting small error. (I haven't tested this either but I assume the current GTS code is flooring the 20837583.75 to 20837583.
> 
> I am wondering if changing the iio_gts_get_gain() to do rounding instead of flooring and changing also the iio_gts_total_gain_to_scale() to something like:
> 
> int iio_gts_total_gain_to_scale(struct iio_gts *gts, int total_gain,
>                  int *scale_int, int *scale_nano)
> {
>      u64 tmp;
>      int rem;
> 
>      tmp = gts->max_scale;
> 
>      rem = do_div(tmp, total_gain);
>      if (total_gain > 1 && rem >= total_gain / 2)
>          tmp += 1ULL;
> 
>      return iio_gts_delinearize(tmp, NANO, scale_int, scale_nano);
> }
> 
> would do the trick. It's just that I'm a bit afraid of touching the iio_gts_get_gain() - by the very least I need to fire up the GTS tests which I implemented but are not in-tree due to the test-device dependency... :/
> 
> Any thoughts?
> 
Hi Matti,
Sorry, got busy with my full time job.
It's nice to see that you have found the issue without my test results:)

Please find below my tests -

root@stm32mp1:/sys/bus/iio/devices/iio:device1# cat scale_available
14.009712000 4.669904000 2.334952000 1.751214000 1.556634666 0.875607000 0.778317333 0.583738000 0.437803500 0.291869000 0.218901750 0.194579333 0.145934500 0.109450875 0.097289666 0.072967250 0.048644833 0.036483625
0.024322416 0.018241812 0.012161208 0.006080604
root@stm32mp1:/sys/bus/iio/devices/iio:device1# echo 0.875607000 > scale ## This works
root@stm32mp1:/sys/bus/iio/devices/iio:device1# echo 0.097289666 > scale ## This fails
root@stm32mp1:/sys/bus/iio/devices/iio:device1# echo 0.097289667 > scale ## However if I add 1, it works! I figured, its a rounding issue so used this trick: "if (val2 % 10) val2 += 1;"
I am sorry, I haven't gone through the full gts internals and only used your driver as a reference to understand it's implementation. I do not have any thoughts on top of my head now but let me go through the code.
   
Regards,
Subhajit Ghosh