On 20 July 2016 at 14:37, Quentin Schulz
[off-list ref] wrote:

On 18/07/2016 15:18, Jonathan Cameron wrote:
[...]

quoted

quoted

+
+    if (!wait_for_completion_timeout(&info->completion,
+                                     msecs_to_jiffies(100))) {
+            ret = -ETIMEDOUT;
+            goto out;
+    }
+
+    if (info->flags & SUNXI_GPADC_ARCH_SUN4I)
+            *val = info->temp_data * 133 - 257000;

Why report as processed?  I'd just report them as raw with the scale
and offset provided.  It's not a big thing, but if we can leave it so
that the conversion only occurs when desired, why not?

For in kernel users, this all happen 'automagically' anyway ;)

Mmmmh, in the code above we apply the scale on the raw value and then
the offset. While in iio_convert_raw_to_processed
(http://lxr.free-electrons.com/source/drivers/iio/inkern.c#L507), the
offset is applied before the scale.

The way would be to factorize the computation by scale:
Now: *val = raw * scale + offset
Then: *val = (raw + offset/scale) * scale

But the offset is an integer and offset/scale is therefore rounded.
Currently, we have the following values:
sun4i: -257000/133 = -1932.3308270676691
sun5i: -144700/100 = -1447
sun6i: -271000/167 = -1622.754491017964

Do we accept such rounding?

How much does this rounding bring as mistake on end result and is this
acceptable for you? You can always multiply it by 1000 or more to get
the precision you want out.

If not, we either stay with the processed value in read_raw or patch
inkern to add an offset to apply after having applied the scale to the
raw value (val2 from iio_channel_read is yet unused with
IIO_CHAN_INFO_OFFSET for example, we could use that to specify an
offset2 to apply after the switch(scale_type)-case).

[...]
Quentin