[Pixman] [PATCH 01/11] mmx: add store function and use it in add_8888_8888

[Søren Sandmann]

Matt Turner <mattst88 at gmail.com> writes:


>> - add_8_8: Here too you also only do the optimization in the edge
>>  cases.
>
> I can't (at least with lowlevel-blt-bench) measure any performance
> improvements by doing the optimization in the middle of scanlines. I
> suppose it's because the optimization would only skip a block of 4
> pixels when all are all zero. That might be an uncommon case in a
> synthetic benchmark like lowlevel-blt-bench. Another thing is that
> you've got to wait until the src is loaded before you can know whether
> you need to load dst, which potentially doubles the amount of time
> you're waiting for memory.
>
> I think because of the extra memory latency it might not be worth
> doing the optimization.
>
> On the other hand, as you can see it does help when doing the
> optimization on the edge cases.

Is there a reasonable explanation that it would be a win for edge cases
and not for the middle cases?

If you really think this patch is an improvement, I probably won't
complain too much since you are the one with the hardware and the one
doing the work, but I have to say that data dependent optimizations that
are only shown to improve lowlevel-blt and only in particular special
cases, feel a bit too much like voodoo for my taste.

> I'll plan to try to find some cairo traces that exercise the in_*
> functions. I doesn't look like they are very important, since we don't
> have NEON fast paths for them! :)

They used to be very important for Evince, but it may be that newer
versions of cairo end up taking different paths.

> With the in_8_8 patch dropped and the other comments addressed, good to commit?

I guess.


Thanks,
Soren