[Pixman] performance of bilinear_interpolation

[Maksim Lin]

HI Siarhei,

Thanks for taking a look at this.
I'm away on leave so I didn't get a chance to reply to Søren earlier
and say thanks to him having a look too.

The thing is though, I see the poor performance for this webpage
happening when we are displaying this on a 1920x1080 monitor (hence
the size of the images in that slideshow) and it seems the content
fits the screen (we customised android to make firefox be fullscreen
with no chrome) but perhaps we've done something wrong there and
firefox is zooming the content so I'll follow that up straight away to
double check. Though because the firefox profiler add-on doesnt
support native stack traces I can't be sure the on x86 we are seeing
exactly the same slow-path issue as when profiling with arm devies.

On a couple of arm devices (Nexus10 and a Amlogic AML8726-M based
tablet) where I can get full traces from the  firefox profiler add-on
I do see which pixman functions are being called and see the slow path
being use a lot of the time.

Now here is the strange thing: on another Arm tablet, a Allwinner A10
based one (same res 1024x768 as the Amlogic tablet) I *dont* see the
slow path being taken, the Neon based function are always called. The
rendering performance is still visually not that great on the A10
tablet, but it is roughly the same as for the webkit based default
browser, which on all the other devices is drastically better then
firefox.

I haven't come across how to do cairo traces, would these work on x86
based android devices when running firefox? If so I'd be very happy to
run them and see if I can find more useful information, would you be
able to point me to some instructions on how to run them?

thanks again,
Maks.

On Thu, Sep 26, 2013 at 9:20 AM, Siarhei Siamashka
<siarhei.siamashka at gmail.com> wrote:
> On Wed, 11 Sep 2013 17:07:52 +1000
> Maksim Lin <maks at manichord.com> wrote:
>
>> Hi,
>>
>> I'm working on finding out the cause of poor drawing performance of
>> Firefox on android:
>> https://bugzilla.mozilla.org/show_bug.cgi?id=911882
>>
>> The html causing the problem is essentially doing a image translation:
>> http://manichord.com/opensign/ffdebug/ffscroll.html
>
> When this web page is zoomed in Firefox, we get bilinear scaling for
> x8r8g8b8 -> a8r8g8b8 (SRC operator, PAD repeat). And there is simply no
> NEON or SSE2 optimized fast path in the currently released versions of
> pixman (or in the Mozilla's internal pixman copy), which could do this
> scaling while also setting the alpha channel to 0xFF at the same time.
>
> A universal general solution would be to extend the SSSE3 bilinear
> fetch iterator in pixman git repository to handle all repeat types
> and also add NEON and SSE2 bilinear fetch iterators.
>
> But first it would be a good idea to record a cairo trace for this
> use case.
>
> --
> Best regards,
> Siarhei Siamashka