On Wed, Sep 27, 2023 at 03:17:06PM +0300, Andy Shevchenko wrote:

On Wed, Sep 27, 2023 at 09:32:11AM +0800, Kent Gibson wrote:

quoted

On Tue, Sep 26, 2023 at 08:20:07AM +0300, Andy Shevchenko wrote:

quoted

Currently we have a few bitmap calls that are open coded in the library
module. Let's convert them to use generic bitmap APIs instead.

Firstly, I didn't consider using the bitmap module here as, in my mind at
least, that is intended for bitmaps wider than 64 bits, or with variable
width. In this case the bitmap is fixed at 64 bits, so bitops seemed more
appropriate.

And I would argue that they aren't "open coded" - they are parallelized
to reduce the number of passes over the bitmap.
This change serialises them, e.g. the get used to require 2 passes over
the bitmap, it now requires 3 or 4.  The set used to require 1 and now
requires 2.
And there are additional copies that the original doesn't require.
So your change looks less efficient to me - unless there is direct
hardware support for bitmap ops??

Wrt the argument that the serialized form is clearer and more
maintainable, optimised code is frequently more cryptic - as noted in
bitmap.c itself, and this code has remained unchanged since it was merged
3 years ago, so the only maintenance it has required is to be more
maintainable??  Ok then.

Your patch is functionally equivalent and pass my uAPI tests, so 

Tested-by: Kent Gibson <warthog618@gmail.com>

Thanks for testing!

Not a problem - that is what test suites are for.

quoted

but my preference is to leave it as is.

As Yury mentioned we need to look at bitmap APIs and make them possible to have
a compile-time optimizations. With that in mind, I would prefer bitmap APIs
over open-coded stuff which is hardly to be understood (yes, I still point
out that it takes a few hours to me, maybe because I'm stupid enough, to
get what's the heck is going one there, esp. for the == 1 case).

Really?  With all the bits out in the open it seems pretty clear to me.
Clearer than scatter/gather in fact.

Sure, if there is suitable hardware support then bitmaps COULD be faster
than bitops.  But without that, and that is the general case, it will be
slower.  Do you have ANY cases where your implementation is currently
faster?  Then you would have a stronger case.

And if you find the existing implementation unclear then the appropriate
solution is to better document it, as bitmaps itself does, not replace it
with something simpler and slower.

Yet, it opens a way to scale this in case we might have v3 ABI that let's say
allows to work with 512 GPIOs at a time. With your code it will be much harder
to achieve and see what you wrote about maintenance (in that case).

v3 ABI?? libgpiod v2 is barely out the door!
Do you have any cases where 64 lines per request is limiting?
If that sort of speculation isn't premature optimisation then I don't know
what is.

Cheers,
Kent.

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel