On 9/6/24 12:34 AM, Alexandru Ardelean wrote:
> On Fri, Sep 6, 2024 at 2:30 AM David Lechner <dlechner@xxxxxxxxxxxx> wrote:
>>
>> On 9/5/24 3:24 AM, Alexandru Ardelean wrote:
>>> -static int ad7606_read_samples(struct ad7606_state *st)
>>> +static int ad7606_read_samples(struct ad7606_state *st, bool sign_extend_samples)
>>> {
>>> + unsigned int storagebits = st->chip_info->channels[1].scan_type.storagebits;
>>
>> Why [1]? Sure, they are all the same, but [0] would seem less arbitrary.
>
> [0] is the timestamp channel.
Oh, that's weird. First channel but last scan index!?
>>
>>> + if (ret)
>>> + return ret;
>>> +
>>> + if (storagebits == 16 || !sign_extend_samples)
>>> + return 0;
>>> +
>>> + /* For 18 bit samples, we need to sign-extend samples to 32 bits */
>>> + for (i = 0; i < num; i++)
>>> + data32[i] = sign_extend32(data32[i], 17);> +
>>> + return 0;
>>> }
>>>
>>> static irqreturn_t ad7606_trigger_handler(int irq, void *p)
>>> @@ -124,11 +176,11 @@ static irqreturn_t ad7606_trigger_handler(int irq, void *p)
>>>
>>> guard(mutex)(&st->lock);
>>>
>>> - ret = ad7606_read_samples(st);
>>> + ret = ad7606_read_samples(st, true);
>>
>> Shouldn't the sign_extend parameter depend on if the data is unipolar or bipolar?
>
> [c1]
> Sign-extension is only needed for 18-bit samples.
> 16-bit samples are already properly sign(ed), but to 16-bits.
>
> It's a slight performance improvement, that may look quirky here.
> The idea here, is that for ad7606_scan_direct() we only need to
> sign-extend 1 sample of the 8 samples we get.
> And we need to sign-extend it to 32 bits regardless of it being 16-bit
> or 18-bit.
>
> In ad7606_trigger_handler(), the 16-bit samples were pushed as-is.
> Which means that we need to sign-extend the samples at least for
> 18-bits (as it is a new part)
> The question now becomes if we should sign-extend to 32-bits, 16-bit
> samples in ad7606_trigger_handler(), as that may break some ABI.
>
Sign extension should not be needed at all for buffered reads (that is
what scan_type is for). So sign extension should only be needed for
the direct read when returning a raw value via sysfs (raw read).
