On Sat Sep 14, 2024 at 11:25 AM UTC, Jonathan Cameron wrote:
> On Fri, 13 Sep 2024 15:46:17 +0200
> Nuno Sá <noname.nuno@xxxxxxxxx> wrote:
>
> > On Fri, 2024-09-13 at 12:55 +0000, Esteban Blanc wrote:
> > > On Fri Sep 13, 2024 at 10:18 AM UTC, Nuno Sá wrote:
> > > > On Fri, 2024-09-13 at 09:55 +0000, Esteban Blanc wrote:
> > > > > On Mon Aug 26, 2024 at 9:27 AM UTC, Jonathan Cameron wrote:
> > > > > > On Thu, 22 Aug 2024 14:45:20 +0200
> > > > > > Esteban Blanc <eblanc@xxxxxxxxxxxx> wrote:
> > > > > > > +static const unsigned long ad4630_channel_masks[] = {
> > > > > > > + /* Differential only */
> > > > > > > + BIT(0) | BIT(2),
> > > > > > > + /* Differential with common byte */
> > > > > > > + GENMASK(3, 0),
> > > > > > The packing of data isn't going to be good. How bad to shuffle
> > > > > > to put the two small channels next to each other?
> > > > > > Seems like it means you will want to combine your deinterleave
> > > > > > and channel specific handling above, which is a bit fiddly but
> > > > > > not much worse than current code.
> > > > >
> > > > > I can do it since that was what I had done in the RFC in the first place.
> > > > > Nuno asked for in this email
> > > > > https://lore.kernel.org/r/0036d44542f8cf45c91c867f0ddd7b45d1904d6b.camel@xxxxxxxxx/
> > > > > :
> > > > >
> > > > > > > > * You're pushing the CM channels into the end. So when we a 2 channel
> > > > > > > > device
> > > > > > > > we'll have:
> > > > >
> > > > > > > > in_voltage0 - diff
> > > > > > > > in_voltage1 - diff
> > > > > > > > in_voltage2 - CM associated with chan0
> > > > > > > > in_voltage0 - CM associated with chan1
> > > > > > > >
> > > > > > > > I think we could make it so the CM channel comes right after the channel
> > > > > > > > where
> > > > > > > > it's data belongs too. So for example, odd channels would be CM channels
> > > > > > > > (and
> > > > > > > > labels could also make sense).
> > > > >
> > > > > So that's what I did here :D
> > > > >
> > > > > For the software side off things here it doesn't change a lot of things
> > > > > since we have to manipulate the data anyway, putting the extra byte at the
> > > > > end or in between is no extra work.
> > > > > For the offload engine however, it should be easier to ask for 24 bits
> > > > > then 8 bits for each channel as it would return two u32 per "hardware
> > > > > channel".
> > > > >
> > > > > In order to avoid having two different layouts, I was kind of sold by
> > > > > Nuno's idea of having the CM in between each diff channel.
> > > > >
> > > >
> > > > Tbh, I was not even thinking about the layout when I proposed the arrangement.
> > > > Just
> > > > made sense to me (from a logical point of view) to have them together as they
> > > > relate
> > > > to the same physical channel. FWIW, we're also speaking bytes in here so not sure
> > > > if
> > > > it's that important (or bad).
> > >
> > > The best we can do (if we managed to do it HDL wise) is to reorder the
> > > data to get both CM byte in a single u32 after the 2 u32 of both diff
> > > channel. That would be 3 u32 instead of 4.
>
> Entirely up to you. :)
Ok so here is the plan I propose:
1. Use the layout of this patch (common byte channels just after their
respective diff channel) as it should work out of the box for the offload
engine (once it's merged [1]).
2. In case of performance issue, switch to the RFC layout (both diff
channels then both common byte channels) and try to modify the HDL for
the offload engine to reduce the memory footprint by one byte for the 2
hardware channels case.
[1]: https://lore.kernel.org/lkml/20240722-dlech-mainline-spi-engine-offload-2-v3-0-7420e45df69b@xxxxxxxxxxxx/
Best regards,
--
Esteban Blanc
BayLibre
