Re: Integer pixel operations for gtk+?

[Rob Taylor <rob taylor codethink co uk>]

Koen Kooi wrote:
> Daniel Amelang schreef:
>>> On 2/10/07, Koen Kooi <koen dominion kabel utwente nl> wrote:
>>>> -----BEGIN PGP SIGNED MESSAGE-----
>>>> Hash: SHA1
>>>>
>>>> Michael 'Mickey' Lauer schreef:
>>>>> Thanks for raising this issue Koen,
>>>>>
>>>>> I'd be very grateful for someone updating this patch to
>>>>>
>>>>> a) 2.6.10 and
>>>>> b) 2.10.x
>>>> Xan Lopez has expressed some interest in updating that patch (probably
>>>> for 2.10.x), but
>>>> won't have time till next week to do proper testing.
>>>>
>>>> There was some talk on the cairo list about moving image stuff into
>>>> cairo and having
>>>> optimized implementations there. It would be a nice experiment since
>>>> the cairo developers
>>>> are a lot more open to suggestions and contributions than the gtk+
>>>> people.
>>> Well it sure makes sense to move "image stuff" into cairo. And we're
>>> always up for a challenge to optimize for a given platform, especially
>>> if you can make a cairo-perf test out of it, or give us some simple
>>> example code that we could turn into a perf test.
>>>
>>> First though, I personally would like to see some actual profiles that
>>> identify this as a bottleneck. To quote Koen from the bugzilla entry:
>>> "Let's hope it doesn't take another 4 and a half year to (produce a
>>> profile)." Federico asked for one a long time ago.
> 
> In the span of the bug in question, a short time ago.
> 
>>> Maybe that's why
>>> the bug hasn't been recieved the needed love?
> 
> Point me to some code to do some performance test on and/or profile and I happily do that.
> I asked for that before....
> 

Koen, TBH its pretty easy to figure out a performance test for this
patch; Write a test that loads in a test image with GdkPixbufLoader, and
then scales it to several different resolutions, repeating each scale
some number of times (say, 100), and exits when finished. Then just time
that test, running it a number of times and calculate the average (and
std deviation, for extra bonus points). Run that against vanilla and
patched (and maybe also a patched version with the Cairo fast
double->fixed method). Of course, it'd also be useful for testing a
version that just calls out to cairo.

If you can get a test up and running, I'll happily run it on N800 as
well for comparison.

Thanks,
Rob Taylor