> > If you are using bursts, the data is getting read anyway - which is the main > cost here. The real question becomes what are you actually saving by supporting all > the combinations of the the two subsets of channels in the pollfunc? > Currently you have to pick the channels out and repack them, if pushing them all > looks to me like a mempcy and a single value being set (unconditionally). I did not get a chance to look at this again until now. Unfortunately, a memcpy will not work. The current implementation is as follows: /* The lower register data is sequenced first */ st->data[i++] = buffer[2 * bit + offset + 3]; st->data[i++] = buffer[2 * bit + offset + 2]; The device first sends the 16LSB, then the next 16MSB in big endian format. So then I wonder, can we keep the same implementation logic? The code is implemented in the same manner for adis16475 driver which uses the same channels data packing approach. > > Then it's a question of what the overhead of the channel demux in the core is. > What you pass out of the driver via iio_push_to_buffers*() > is not what ends up in the buffer if you allow the IIO core to do data demuxing > for you - that is enabled by providing available_scan_masks. At buffer > start up the demux code computes a fairly optimal set of copies to repack > the incoming data to match with what channels the consumer (here probably > the kfifo on the way to userspace) is expecting. > > That demux adds a small overhead but it should be small as long > as the channels wanted aren't pathological (i.e. every other one). > > Advantage is the driver ends up simpler and in the common case of turn > on all the channels (why else did you buy a device with those measurements > if you didn't want them!) the demux is zerocopy so effectively free which > is not going to be the case for the bitmap walk and element copy in the > driver. > > Jonathan >
